Journal of Memory and Language 56 (2007) 384–409
Journal of
Memory and
Language
www.elsevier.com/locate/jml
The effect of syntactic constraints on the processing
of backwards anaphora
Nina Kazanina a,*, Ellen F. Lau b, Moti Lieberman c,
Masaya Yoshida b, Colin Phillips b,d
a
Department of Linguistics, University of Ottawa, 401 Arts Hall, 70 Laurier Avenue East, Ottawa, Ont., KIN 6N5, Canada
b
Department of Linguistics, University of Maryland, College Park, MD 20742, USA
c
Department of Linguistics, McGill University, 1085 Dr. Penfield, Montreal, QC H3A 1A7, Canada
d
Neuroscience and Cognitive Science Program, University of Maryland, College Park, MD 20742, USA
Received 5 February 2006; revision received 20 September 2006
Available online 3 November 2006
Abstract
This article presents three studies that investigate when syntactic constraints become available during the processing
of long-distance backwards pronominal dependencies (backwards anaphora or cataphora). Earlier work demonstrated
that in such structures the parser initiates an active search for an antecedent for a pronoun, leading to gender mismatch
effects in cases where a noun phrase in a potential antecedent position mismatches the gender of the pronoun [Van
Gompel, R. P. G. & Liversedge, S. P. (2003). The influence of morphological information on cataphoric pronoun
assignment. Journal of Experimental Psychology: Learning, Memory, and Cognition, 29, 128–139]. Results from three
self-paced reading studies suggest that structural constraints on coreference, in particular Principle C of the Binding
Theory [Chomsky, N. (1981). Lectures on government and binding. Dordrecht, Foris], exert an influence at an early stage
of this search process, such that gender mismatch effects are elicited at grammatically licit antecedent positions, but not
at grammatically illicit antecedent positions. The results also show that the distribution of gender mismatch effects is
unlikely to be due to differences in the predictability of different potential antecedents. These findings suggest that backwards anaphora dependencies are processed with a grammatically constrained active search mechanism, similar to the
mechanism used to process another type of long-distance dependency, the wh dependency (e.g., [Stowe, L. (1986). Evidence for online gap creation. Language and Cognitive Processes, 1, 227–245; Traxler, M. J., & Pickering, M. J. (1996).
Plausibility and the processing of unbounded dependencies: an eye-tracking study. Journal of Memory and Language,
35, 454–475.]). We suggest that the temporal priority for syntactic information observed here reflects the predictability
of structural information, rather than the need for an architectural constraint that delays the use of non-syntactic
information.
2006 Elsevier Inc. All rights reserved.
Keywords: Parsing; Coreference; Binding; Cataphora; Grammatical constraints; Pronouns
*
Corresponding author. Fax: +1 613 562 5141.
E-mail address: [email protected] (N. Kazanina).
0749-596X/$ - see front matter 2006 Elsevier Inc. All rights reserved.
doi:10.1016/j.jml.2006.09.003
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
Introduction
To understand the architecture of the human language
processor it is important to establish the extent to which
its behavior is governed by general mechanisms or by
highly specific subroutines that are restricted to individual
constructions or individual languages. In this article we
address the question of the generality of sentence processing mechanisms by examining whether well-known properties of the processing of long-distance filler-gap
dependencies are also found in the processing of backwards pronominal dependencies (backwards anaphora or
cataphora), as found in sentences like When she was in
Paris, Susan visited the Louvre. Backwards anaphoric
dependencies parallel filler-gap dependencies in terms of
the relative ordering of key elements, but differ in a number of other respects, including the obligatoriness and
overtness of the elements of the dependency and the syntactic and discourse constraints that they are subject to.
An improved understanding of how backwards anaphora
is processed may contribute to a fuller understanding of
the mechanisms responsible for the construction of
long-distance relations in language in general.
Much previous research on filler-gap dependencies
such as wh interrogatives, topicalizations and scrambling
constructions has established two generalizations that
form the starting point for the present studies. First,
much evidence indicates that the parser constructs filler-gap dependencies actively, i.e., after encountering a
suitable filler (e.g., a wh-phrase) it constructs gap positions without waiting for unambiguous evidence for
the position of the gap in the input (e.g., Crain & Fodor,
1985; Frazier & Clifton, 1989; Stowe, 1986). Second, the
mechanisms that search for potential gap positions show
on-line sensitivity to grammatical constraints on fillergap dependencies, such that active gap construction is
not observed in structural positions that are ruled out
by those constraints (e.g., Stowe, 1986; Traxler & Pickering, 1996). In this study we investigate whether these
two properties are also found in the processing of backwards anaphora. Previous studies of the processing of
backwards anaphora provide some evidence for active
dependency formation mechanisms (Van Gompel &
Liversedge, 2003) and for the impact of grammatical
constraints (Cowart & Cairns, 1987), but in both cases
the findings are open to alternative interpretations. In
the current study we examine these issues in more detail,
with a particular focus on distinguishing the effects of
grammatical constraints from the effects of distributional
probabilities on active dependency formation.
Properties of long-distance dependencies
The wh-constructions in (1) are examples of a longdistance dependency that create a relation between a
385
wh-phrase, also known as a filler, and its gap position,
as in (1a). Although the filler and the gap may potentially be separated by an indefinite amount of intervening material (1b–c), the occurrence of the filler reliably
predicts the occurrence of a gap within the same sentence and if there is no gap, the result is unacceptable
(1d). Note that we use the filler-gap terminology here
mostly for the purposes of exposition. The long-standing controversy about whether wh-dependencies genuinely involve gaps or whether they involve direct
associations between verbs and wh-phrases (for reviews
see Phillips & Wagers, in press; Pickering, 1993) is
orthogonal to the processing issues that are the focus
of this article.
(1)
a. What did John see __?
b. What did Mary say that John saw __?
c. What did Bill say that Mary claimed that John
saw __?
d. *What did John see an apple?
The relative positioning of the two elements of a whdependency is subject to a number of constraints on
well-formedness, often known as island constraints.
The sentences in (2), for example, are unacceptable
due to a ban on filler-gap dependencies that cross relative clause boundaries as in (2a) or that span non-complement adverbial clauses as in (2b) (Chomsky, 1973;
Huang, 1982; Ross, 1967).
(2) a. *What did the boy that bought __ like
Mary?
b. *What did the boy read a book while
his sister was playing with __?
Referential dependencies between a pronoun and
an antecedent noun phrase include forwards anaphora, in which the antecedent precedes the pronoun
(3a), and backwards anaphora, in which the pronoun
precedes the antecedent (3b). Similar to filler-gap
dependencies, referential dependencies can span multiple clauses, as in (3a), and are subject to structural
restrictions. Most relevant for the current study is a
constraint, Principle C (Chomsky, 1981), that prohibits referential dependencies in which the pronoun
structurally c-commands the antecedent, i.e., where
the antecedent is structurally within the scope of
the pronoun (4).
(3) a. Johni said that the news report would
make the customers think that hei wanted
to sell the company.
b. Although shei was sleepy, Susani tried
to pay attention.
386
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
(4) *Hei said that Johni saw the movie.
intended meaning: ‘John said that he
(=John) saw the movie.’
In research on language processing, the constraints
on long-distance filler-gap dependencies have received
a good deal of attention, but there has been relatively little research on the processing of constraints on backwards anaphora.
Real time processing of wh-dependencies
Fodor (1978) sketched out two potential mechanisms
for parsing filler-gap dependencies, contrasting a ‘gapdriven’ view whereby dependency formation is initiated
only when the parser has encountered all necessary pieces of information in the input, that is, both the filler and
the gap, with a ‘filler-driven’ view, according to which
the parser starts predictively building a dependency
upon encountering the filler that constitutes the first half
of the dependency. Stowe (1986) explored this distinction using a word-by-word self-paced reading technique,
by comparing reading times for sentences that contain a
licit wh-dependency (5a) with controls that lack a whdependency (5b).
(5) a. My brother wanted to know
who Ruth will bring us home to __ at
Christmas.
b. My brother wanted to know if Ruth
will bring us home to Mom at Christmas.
Stowe found that readers slowed down at the direct
object noun phrase us in wh-sentences, relative to the
corresponding word in the control sentence, and
argued that this effect arose because the parser predictively posited a gap in direct object position, which
subsequently needed to be retracted upon encountering
the overt pronoun us. This ‘filled-gap effect’ is taken to
support the ‘filler-driven’ view and to reflect the parser’s willingness to complete a wh-dependency before it
has unambiguous bottom-up evidence that the verb
has a missing argument. Such a mechanism for fillergap dependency formation was dubbed as active
search, and has been supported by evidence from similar filled-gap effects in diverse syntactic positions
(Aoshima, Phillips, & Weinberg, 2004; Crain & Fodor,
1985; Lee, 2004), and by evidence using different
experimental measures, including eye-tracking (Sussman & Sedivy, 2003; Traxler & Pickering, 1996),
event-related potentials (Garnsey, Tanenhaus, &
Chapman, 1989; Kaan, Harris, Gibson, & Holcomb,
2000), and cross-modal lexical priming (Nicol, Fodor,
& Swinney, 1994).
The evidence for active wh-dependency formation
leads to a further question about the grammatical accuracy of the parser. Whereas a gap-driven parsing mechanism could be expected to posit gaps only in
grammatically acceptable positions, by virtue of its
dependence on direct evidence in the language input, this
is not guaranteed for an active dependency formation
mechanism, since it is predictive and less closely tied to
evidence in the bottom-up input. If the process of active
dependency formation immediately adheres to grammatical constraints, the parser should not attempt to
posit a gap inside an island, despite its preference to
complete the dependency as soon as possible. In her second experiment, Stowe (1986) found that no filled-gap
effect was observed inside a syntactic island, suggesting
that the parser respects constraints on wh-movement even
at the earliest stages of filler-gap dependency formation
(see also Bourdages, 1992; Phillips, 2006; Traxler &
Pickering, 1996 for a similar conclusion and Pickering,
Barton, & Shillcock, 1994 for discussion of potentially
discrepant findings).
Referential dependencies
Backward anaphoric dependencies parallel fillergap dependencies in the respect that they may span
long-distances, and in the respect that a dependent
element (wh-phrase, pronoun) precedes a controlling
element (verb or gap, antecedent noun phrase). This
raises the possibility that the parser may analyze
backward anaphoric dependencies in the same fashion
as filler-gap dependencies, searching for an antecedent
for an unanchored pronoun in a manner that is active
yet sensitive to constraints on referential dependencies. However, unlike wh-fillers that require a gap
position, finding an overt antecedent is not a grammatical requirement for a pronoun, which can be taken to refer to an unspecified discourse referent. For
this reason, an active search mechanism may not be
as useful for backwards anaphora as for filler-gap
dependencies. Furthermore, if common mechanisms
are used for parsing forwards and backwards referential dependencies, then this is another reason why
active search might not be deployed, since forward
anaphoric dependencies cannot be recognized until
both the antecedent and the pronoun have been
encountered.
If pronouns do trigger an active search for an antecedent, this leads to the question of whether binding
principles immediately constrain active search for antecedents. If binding principles, like island constraints on
filler-gap dependencies, constrain the earliest stages of
structure generation, we should expect that during an
active search for an antecedent triggered by a cataphoric
pronoun the parser should never consider positions that
would violate the binding constraints. Note that our use
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
of the term ‘antecedent’ in referring to a noun phrase
does not necessarily entail that that noun phrase is a licit
licensor for the dependent element in question. Instead it
will be used as a shortcut for ‘a candidate antecedent’
which could become a licit antecedent provided that (i)
it occurs in a licit structural position, and (ii) it matches
the pronoun in its morphological features (gender,
number).
Active search in backwards anaphora processing
Previous evidence for active search for antecedents in
the processing of backwards anaphora comes from a
study by Van Gompel and Liversedge (2003) that used
eye-tracking to examine reading times for bi-clausal sentences as in (6). The gender of the main subject was
manipulated such that it either matched (6a) or mismatched (6b) with the preceding pronoun. In all sentences an antecedent for the pronoun appeared at some
point in the main clause.
(6) Stimuli from Experiment 1, Van Gompel and
Liversedge (2003).
a. gender match
When he was at the party, the boy cruelly
teased the girl during the party games.
b. gender mismatch
When he was at the party, the girl cruelly
teased the boy during the party games.
c. control
When I was at the party, the boy cruelly
teased the girl during the party games.
Van Gompel and Liversedge found a mismatch
effect in early eye-tracking measures at the region
immediately following the main clause subject noun:
first-pass reading times were slower at cruelly in (6a)
than in (6b). Furthermore, inclusion of the control condition in (6c) made it possible to rule out the possibility
that the longer reading times following the second subject in (6b) were due to the introduction of a new discourse entity, since reading times in the critical region
in the control condition (6c) did not differ from (6a).
Van Gompel and Liversege instead argue that this gender mismatch effect reflects formation of a referential
dependency between the pronoun and the second subject noun phrase before relevant bottom-up information about the antecedent has been taken into
consideration, which leads to processing difficulty when
the antecedent noun phrase is recognized as semantically incompatible with the pronoun. This evidence parallels the filled-gap effect found in filler-gap dependency
processing, and may be considered as evidence for
active formation of backwards anaphoric dependencies
387
(for similar findings from Japanese see Aoshima, 2003;
Aoshima, Yoshida, & Phillips, submitted for
publication).
Note that a number of different processing mechanisms for backward anaphora may be considered as
‘active’, in the respect that dependencies are constructed before all bottom-up information has been analyzed. Van Gompel and Liversedge (2003) assume
that the syntactic category of the antecedent must be
processed before a referential dependency can be
formed, whereas computation of its morphological
properties ‘is delayed until after the computation of
coreference relations’ (p. 128), and thus argue that
syntactic category information has an architectural
priority in parsing (e.g., Cowart & Cairns, 1987;
Fodor, 1983; Frazier, 1987). However, Van Gompel
and Liversedge’s findings are also compatible with
an even more ‘active’ processing mechanism that constructs referential dependencies as soon as an antecedent position can be reliably predicted, potentially
before any of the specific features of the antecedent
are encountered in the input. Under this approach,
the parser may recognize that a sentence-initial
while-clause, as in (6), must be followed by a main
clause that contains a subject noun phrase, and may
use this knowledge to construct a dependency between
the pronoun and the main clause subject even before
the end of the while-clause. If this is the case, then
the gender mismatch effect may reflect the fact that
a subject position may be reliably projected ahead of
time as a structurally licit antecedent, whereas its gender and other features may only become known once
the subject noun phrase is encountered bottom-up.
This view would obviate the need to impose an architectural constraint that delays the availability of morphological information, as Van Gompel and
Liversedge propose.
The gender mismatch paradigm therefore provides a
useful method for tracking active dependency formation
in backwards anaphora, but leaves open the question of
how ‘active’ the search for pronoun antecedents is, and
also how sensitive it is to constraints on referential
dependencies.
Processing of constraints on forwards anaphora
A number of previous studies explored the timecourse of the application of binding constraints on
forwards anaphora constructions in which the pronoun follows the antecedent, in particular the requirement that reflexives find a clausemate antecedent (7)
and the requirement that pronouns not have a
clausemate antecedent (8), known as Principle A and
Principle B respectively (Chomsky, 1981; Reinhart &
Reuland, 1993).
388
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
(7) a. Susan heard that Billi had painted himselfi.
b.*Billi heard that Susan had painted himselfi.
(8) a. *Susan heard that Billi had painted himi.
b. Billi heard that Susan had painted himi.
Evidence has been found for both early/immediate
and late/delayed application of these constraints. Nicol
and Swinney (1989) examined the activation of candidate antecedents at a pronoun or reflexive using a
cross-modal priming task and found that candidates
excluded by Conditions A and B were not considered
(see also Clifton, Kennison, & Albrecht, 1997). In contrast, Badecker and Straub (2002) showed in a selfpaced reading task that reading times for a pronoun
were affected by the gender congruency of a grammatically inaccessible antecedent, and concluded from this
that the application of grammatical constraints on
binding was delayed relative to early parsing processes
(see also Kennison, 2003). More recently, Sturt (2003)
found in eye-tracking studies on the processing of
reflexives that early eye-tracking measures (first-fixation and first-pass) were unaffected by the gender
congruency of grammatically inaccessible antecedents,
although some later eye-tracking measures (secondpass, regression path) did show effects of those
antecedents. Sturt’s findings raise the possibility that
discrepancies among earlier findings might have been
due to differences in the stage of processing that different tasks tapped into. This parallels findings on the
processing of filler-gap dependencies, where evidence
for the construction of illicit dependencies comes
primarily from studies that have used whole sentence
judgment tasks.
A difficulty in interpreting effects of grammatical constraints on the processing of forwards anaphora arises
from the fact that the pronoun or reflexive cannot, in
general, be anticipated, and therefore the search for an
antecedent cannot be initiated until the end of the
dependency is reached. Consequently, the search for
an antecedent involves a search in memory for candidate
antecedents. If there are multiple candidate antecedents,
they may need to be evaluated in parallel. Studies of
binding constraints in processing have asked whether
the parser selectively searches structural positions where
antecedents are allowed, but this question overlaps with
the question of what type of memory representation is
searched. The pieces of a sentence may be encoded in
echoic memory, in a syntactic parse, or in a discourse
model, and may to some degree be stored in all of these
simultaneously. Therefore, evidence for effects of grammatically inaccessible antecedents on the processing of
pronouns or reflexives may reflect processes that search
illegal antecedent positions in a syntactic structure, or
they may reflect processes that access a non-syntactic
encoding of potential antecedents. Furthermore, in cases
where the effect of structurally inaccessible antecedents
is realized as a processing disruption due to the presence
of two candidate antecedents (e.g., Badecker & Straub,
2002; Sturt, 2003), this may reflect interference effects
at a lexical or conceptual level that make a grammatically accessible antecedent harder to retrieve from memory
when it is similar to another noun phrase in the sentence.
These difficulties do not arise in the processing of backwards anaphora, since the dependency can be anticipated after processing of its first element, i.e., the pronoun,
and therefore each potential antecedent can be evaluated
in succession. It is therefore possible to investigate, for a
series of different structural positions, whether the
search for an antecedent for the pronoun is sensitive to
grammatical constraints on backwards anaphora. This
may be viewed as a particularly strong test of the impact
of binding constraints in parsing, since successful constraint application requires that the parser ignore a candidate antecedent at a point where no other antecedent
is yet available.
Principle C in grammar and processing
Although backwards anaphora is productive and fully acceptable in English (see van Hoek, 1997 for multiple
naturalistic examples), it is excluded in configurations
where a pronoun c-commands its antecedent, as in the
examples in (9), where the noun phrases he and John
cannot be understood as coreferential. This constraint,
known as Principle C (Chomsky, 1981), is highly robust
across languages (e.g., Baker, 1991; Jelinek, 1984; but cf.
Speas, 1990; Bruening, 2001) and appears early in language development (Crain & McKee, 1985; Kazanina
& Phillips, 2001).
(9) a. *Hei likes Johni.
b. *Hei said that Johni likes wine.
c. *Hei drank beer while Johni watched a
soccer game.
It should be noted that there are at least two welldefined classes of apparent exceptions to Principle C,
which must be taken into consideration when designing
studies on this constraint. The first type of exception
can be seen in a sentence like He then did what John
always did in such situations, where coreference between
he and John is acceptable. Such cases are generally
understood to involve comparisons of multiple mental
representations or ‘guises’ of the same individual
(e.g., Heim, 1992; Reinhart, 1983), and minimal changes to the situations under comparison can lead to failure of coreference, as in *He then did what John had
done half an hour earlier. Rather than undermine the
structural account of Principle C, such contrasts suggest that we should understand coreference in terms
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
of elements in a mental model, which may stand in a
many-to-one relation with entities in the world. A second type of exception involves sentences like He was
threatening to leave when Billy noticed that the computer
had died (Harris & Bates, 2002). Such cases are clearly
acceptable, and are most effective in descriptions of scenarios where the embedded clause event interrupts the
main clause event. Again, minimal changes make coreference unacceptable, as in *He decided to leave when
Billy noticed that the computer had died. Syntactic tests
suggest, however, that the when-clause in such exceptional examples is a sentential modifier, rather than a
verb phrase-subordinator, which takes away the problematic c-command relation between the pronoun and
the embedded subject and obviates the Principle C violation. In sum, both apparent exceptions to Principle C
may not ultimately be counterexamples to the structural generalization (see Kazanina, 2005 for more detail).
Yet, in all of the studies reported below care was taken
to avoid such exceptions in constructing the experimental materials and the acceptability of all contrasts
under investigation was verified by off-line
questionnaires.
Previous psycholinguistic investigation of Principle
C in sentence processing has been limited. In a
whole-sentence self-paced reading Hirst and Brill
(1980) looked for effects of an inaccessible antecedent
in constructions that contained a potential Principle C
violation using two sentence sequences such as (10),
that contained both a grammatically accessible antecedent for the pronoun (John) and a grammatically
inaccessible one (Henry).
(10) John stood watching. He ran for a doctor after
Henry fell down some stairs.
Hirst and Brill found that reading times for the
second sentence differed as a function of the plausibility of the inaccessible antecedent (Henry) as the referent for the pronoun. They concluded that during
pronoun resolution speakers temporarily consider
grammatically illicit referents. However, the coarse
temporal resolution of the task makes it difficult to
draw inferences regarding the time-course and the
source of the effect. Specifically, the effect could be
due to a later stage of processing at which information
from multiple sources is combined and the sentence is
evaluated for overall plausibility.
Cowart and Cairns (1987) used sentence fragments
such as (11)–(12) to investigate the processing of backwards anaphora and the application of Principle C.
(11) a. Whenever they lecture during the procedure,
charming babies is. . .
b. Whenever you lecture during the procedure,
charming babies is. . .
389
(12) a. If they want to believe that visiting
uncles is. . .
b. If you want to believe that visiting
uncles is. . .
The first phrase of the 2nd clause (charming
babies,visiting uncles), which is ambiguous between a
noun phrase and a gerund, is disambiguated by the number marking on the following auxiliary in favor of the
gerund parse. Cowart and Cairns investigated whether
this ambiguity was affected by a pronoun subject in a
preceding clause. If the pronoun initiates a search for
an antecedent then this may increase the likelihood that
the ambiguous phrase should be analyzed as a noun
phrase in (11a)–(12a), relative to (11b)–(12b), which
contain an indexical pronoun. In either condition there
was an additional restriction against treating the ambiguous phrase as the antecedent of the pronoun: in (11a) it
is a semantically implausible antecedent for the
pronoun; in (12a) it is a grammatically inaccessible antecedent, due to Principle C. In a naming task, Cowart
and Cairns found longer response latencies for is in
(11a) relative to (11b), whereas no difference was found
in (12a) vs. (12b). They concluded that the parser was
more biased towards the noun phrase analysis when it
needed to link the pronoun with an antecedent, but only
if that antecedent was in a grammatically accessible
position. Further, they interpreted the contrast between
(11) and (12) to suggest that semantic restrictions, unlike
syntactic restrictions, do not have the immediate effect of
excluding an incongruent noun phrase from the set of
candidate antecedents for the pronoun.
The findings in Cowart and Cairns (1987) suggest
two important conclusions. First, the cataphoric pronoun they initiates an active search for an antecedent
in the main subject position. In fact, these results
may provide stronger evidence for active search than
Van Gompel and Liversedge’s (2003) study: if the
presence of the pronoun biases the parser to analyze
the ambiguous phrase as a noun phrase, this suggests
that dependency formation does not wait until after
grammatical categories have been identified based
on bottom-up analysis of the input. Secondly, their
results suggest that the parser does not search for
an antecedent in positions that are subject to Principle C. However, some objections can be raised with
respect to the design of the study and this interpretation of the results. First, the naming task that was
used did not provide information on positions
beyond the probe position, which may have concealed important effects at earlier or later words.
Second, as the authors themselves point out, the
results were only significant in the participants analysis but not in the items analysis, which raises the possibility that the results are dependent on the specific
390
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
properties of individual items. Finally, the argument
for immediate effects of Principle C is based on the
fact that the pronoun-type effect is present in the ‘semantic’ pair (11) but not in the ‘syntactic’ pair (12).
This argument requires the implicit assumption that
different semantic and syntactic violations should produce an effect of a comparable magnitude at the
same point in time. In light of this, the conclusions
drawn from these findings should be more cautious,
perhaps that the syntactic considerations are taken
into account by the parser earlier than the semantic
considerations. However, the claim that the parser
never violates a syntactic constraint needs stronger
evidence, e.g., using structures as controls that are
not subject to any constraints. In the experiments
that we present below, we used methods and materials designed to resolve these difficulties.
Experiment 1
Experiment 1 consisted of an acceptability rating task
and an on-line self-paced reading task that used a gender
mismatch paradigm to test for the impact of Principle C
on the search for pronoun antecedents.
Participants
Participants were 60 native speakers of English from
the University of Maryland undergraduate population
with normal or corrected-to-normal vision and no history
of language disorders. All participants in this and subsequent studies gave informed consent and were paid $10/
h for their participation.
Materials and design
The materials for the on-line task are described first,
since the materials for the acceptability rating study were
derived from them. The experiment had five conditions,
four of which were organized in a 2 · 2 design with the
factors constraint(Principle C vs. no-constraint) and gender congruency (match vs. mismatch between the pronoun and the subject of the second clause). A full set of
items is shown in Table 1. The gender of the cataphoric
pronoun was balanced across stimulus sets: half of the
sets were built on the basis of the masculine pronoun
he, and the other half on the basis of the feminine pronoun she. The gender-match and gender-mismatch sentences differed only in the gender of the subject of the
second clause, which was always a gender-unambiguous
proper name, matched for the number of letters and syllables in the match and mismatch variants within each set
of items. To guard against a semantic bias for or against
a coreferential interpretation of the pronoun and the second subject, the two clauses were carefully selected such
that the events in each clause could plausibly be performed either by the same agent or by different agents.
In addition, to ensure that the cataphoric pronoun
received a grammatical antecedent in every case, the target structures were embedded in a further sentence introduced by the conjunctions although or since. The gender
of the third clause subject was chosen such that each sentence had a unique grammatical antecedent for the pronoun. Thus, in the Principle C conditions the subject of
the third clause always matched the gender of the pronoun and served as a grammatical antecedent. In the
no-constraint conditions the gender of the third clause
subject mismatched the pronoun in the gender-match
condition, due to the possibility of coreference between
the pronoun and the second clause subject, but the third
clause subject matched the gender of the pronoun in the
gender-mismatch condition.
Following Van Gompel and Liversedge (2003), we
added a fifth ‘name’ condition to each set. This
condition was identical to the no-constraint/mismatch
condition in the number of referents introduced in each
clause. This was done to ensure that any observed
Table 1
Sample set of experimental items for on-line Experiment 1
Principle C/match
Because last semester shei was taking classes full-time while Kathryn was working two jobs to pay the
bills, Ericai felt guilty
Principle C/mismatch
Because last semester shei was taking classes full-time while Russell was working two jobs to pay the
bills, Ericai felt guilty
No constraint/match
Because last semester while shei was taking classes full-time Kathryni was working two jobs to pay the
bills, Russell never got to see her
No constraint/mismatch
Because last semester while shei was taking classes full-time Russell was working two jobs to pay the
bills, Ericai promised to work part-time in the future
No constraint/name
Because last semester while Ericai was taking classes full-time Russell was working two jobs to pay the
bills, shei promised to work part-time in the future
The underlined name indicates the critical second subject noun phrase. Subscript indices indicate intended backward anaphoric
dependencies.
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
mismatch effect at the second subject in the no-constraint pair could not merely be due to the introduction
of a new discourse referent. If the number of referents is
indeed the reason for a mismatch effect in the no-constraint pair, we should expect a similar increase in reading times in the name condition.
Thirty sets of five conditions were distributed among 5
lists in a Latin Square design, and combined with 90 filler
sentences. To mask experimental sentences, the fillers
bore a number of similarities with the target items, including length and average clause number and were designed
in several subgroups, each built around a salient feature
of the targets, such as the use of proper names and
pronouns, or a subordinator followed by a temporal modifier at the beginning of the sentence. There were no
instances of unresolved anaphora in the filler sentences.
Thus, we ensured that throughout the experiment pronouns always found intra-sentential antecedents. This
was appropriate, since the interest of the study is not in
whether readers search for an antecedent but, rather, in
where they search.
Acceptability rating task (off-line)
We conducted an off-line rating task using a similar
methodology to Gordon and Hendrick (1997). In each
sentence a pronoun and a noun phrase were highlighted in bold and participants were instructed ‘to determine how plausible it is that the pronoun in bold
and the noun in bold refer to the same person’ on a
scale from 1 (impossible) to 5 (absolutely natural).
We sought to confirm that, provided matching
gender/number values, participants indeed accept coreference between the cataphoric pronoun and the second subject in the no-constraint environments, but
reject it in the Principle C environments. 12 sets of
items were chosen from the materials for the online
task of Experiment 1 and then simplified to create
the two conditions shown in (13). Two different lists
were constructed based on these 12 sets using a Latin
Square design, and each list had two different stimulus
order randomizations. 40 native English speakers from
the University of Maryland undergraduate population
were recruited specifically for the purposes of the rating study and completed the questionnaire, which contained 6 instances of each condition, interspersed with
20 additional filler sentences, twelve of which tested
the configurations used in Experiment 2.
(13) Stimuli from off-line rating task,
Experiment 1.
a. * Principle C
Because last semester shei was taking classes
full-time while Kathryn was working
two jobs to pay the bills, Ericai felt guilty.
391
b. No-constraint
Because last semester while shei was taking
classes full-time Kathryni was working two jobs
to pay the bills, Russell never got to see her.
The mean rating score for the Principle C condition
(mean = 1.4, standard error = .12) was significantly
lower than the mean score in the no-constraint condition (mean = 4.1, standard error = .13) both in the participants analysis and items analysis (two-tailed paired
t-test, both ps < .001, t1(39) = 12.2, t2(11) = 38.2).
These results fully agree with the predictions of the
Principle C constraint: coreference between the pronoun in the main subject position and the name in
the second subject position is acceptable only when
the pronoun does not c-command the name, i.e., in
the no-constraint condition but not in the Principle C
condition.
Procedure
Participants were tested using a desktop PC running
the Linger software (Doug Rohde, MIT) in a standard
self-paced word-by-word moving window paradigm
(Just, Carpenter, & Woolley, 1982). Each trial started
with a blank screen. Upon pressing the space bar, a
sentence masked by dashes appeared on the screen.
The masks extended to all letters and punctuation
marks, but left spaces unmarked. As the participant
pressed the spacebar, a new word appeared on the
screen as the previous one was re-masked by dashes.
A comprehension question appeared after the end of
each sentence all at once (e.g. Was Kathryn/Russell
working two jobs?). Participants were instructed to read
sentences at a natural pace and to respond to the comprehension questions as accurately as possible. To
answer the question the subject pressed the f-key for
‘yes’ and the j-key for ‘no.’ If the question was
answered incorrectly the word ‘Incorrect’ appeared
briefly in the center of the screen. Each participant
was randomly assigned to one of the lists, and the
order of the stimuli within the presentation list was
randomized for each participant.
Analysis
Only trials for which the corresponding comprehension question was answered correctly were included in
the analysis. Reading times that exceeded a threshold
of 2.5 standard deviations above a participant’s mean
reading rate for each region were replaced by the
threshold value. This winsorizing procedure affected
2.4% of the data (range 1.9–2.6% for individual
conditions).
392
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
collapsing over participants (F2). Furthermore, min F 0
statistics (Clark, 1973) were computed, and 95%
confidence intervals (CIs) based on the mean squared
errors of the respective effect from the participants
analysis were calculated for comparisons between means
of conditions (Masson & Loftus, 2003).
The regions used for the data analysis corresponded
to single words, except for regions corresponding to
the end of the clause, for which several words were combined due to variation in the clause length between items
(see the legend in Figs. 1 and 2 for regions). The critical
second subject position in which the gender manipulation occurred corresponded to region 8 in all conditions.
The data from the first four conditions were entered
into a 2 · 2 repeated-measures ANOVA with the factors
constraint (Principle C, no constraint) and congruency
(match, mismatch). Reading times from the name condition were compared pairwise to each of the no-constraint conditions in a one-way ANOVA. ANOVAs
were computed on the participant mean raw reading
times collapsing over items (F1), and on item means
Results
Comprehension question accuracy
Two of 60 participants showed comprehension question accuracy that was more than 2.5 standard deviations below the mean accuracy. Consequently, the
lowest scoring participant from each list was removed
from the analysis to balance the number of participants
480
Mean reading time (ms)
Principle C, match
Principle C, mismatch
subject
NP
440
400
360
320
1
2
4
5
6
7
8
9
10
11
12
13
14
15
Because1 / last semester 2 / she4 / was taking5 /classes full-time6 / while7 / Kathryn (Russell)8 /
was working 9 / two10 / jobs to pay the bills,11 / Erica12 / felt13 /guilty14.
Fig. 1. Mean reading times in milliseconds for the Principle C conditions from Experiment 1. The error bars represent one standard
error above/below the participant mean at each region. The arrow marks the critical second subject noun phrase.
Mean reading time (ms)
480
No constraint, match
No constraint, mismatch
Name
440
400
360
subject
NP
320
1
2
3
4
5
6
8
9
10
11
12
13
14
15
Because1 / last semester2 / while3 / she (Erica)4 /was taking5 / classes full-time6 / Kathryn (Russell)8 /
was working 9 / two10 /jobs to pay the bills,11 / Russell (Erica) (she)12 / never13 / got14 / to see her15.
Fig. 2. Mean reading times in milliseconds for the no-constraint and the name conditions from Experiment 1. The error bars represent
one standard error above/below the participant mean at each region. The arrow marks the critical second subject noun phrase.
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
393
Self-paced reading
We first present the results of the 2 · 2 ANOVA on the
Principle C and no-constraint conditions (Table 2), and
then discuss the name condition. The results from the
on each list. For the remaining 55 participants, the mean
accuracy was 91%, with a range of 90–93% for individual conditions. There were no reliable differences in
accuracy between individual conditions.
Table 2
Results of 2 · 2 ANOVAs and min F 0 values for Experiment 1
By participants
Min F 0
By items
df
MSE
F1
p
df
F2
p
df
min F 0
p
Region 4 (1st subject)
Constraint
Congruency
Constraint · congruency
1,54
1,54
1,54
2357
2509
1622
2.0
1.2
2.9
0.160
0.276
0.094
1,29
1,29
1,29
1.6
3.1
1.0
0.216
0.091
0.319
1,70
1,81
1,50
<1
<1
<1
0.348
0.354
0.388
Region 5 (was taking)
Constraint
Congruency
Constraint · congruency
1,54
1,54
1,54
2163
1699
2017
<1
2.0
<1
0.622
0.162
0.768
1,29
1,29
1,29
<1
1.4
<1
0.721
0.244
0.558
1,59
1,66
1,76
<1
<1
<1
0.772
0.366
0.792
Region 6 (classes full-time)
Constraint
Congruency
Constraint · congruency
1,54
1,54
1,54
1785
1227
2300
<1
<1
<1
0.910
0.443
0.946
1,29
1,29
1,29
<1
1.0
<1
0.799
0.320
0.952
1,72
1,83
1,70
<1
<1
<1
0.917
0.541
0.964
Region 8 (2nd subject)
Constraint
Congruency
Constraint · congruency
1,54
1,54
1,54
12602
8026
6982
9.2
<1
4.8
0.004
0.385
0.033
1,29
1,29
1,29
10.9
<1
2.8
0.003
0.419
0.104
1,79
1,72
1,62
5.0
<1
1.8
0.028
0.551
0.188
Region 9 (was working)
Constraint
Congruency
Constraint · congruency
1,54
1,54
1,54
4571
3471
2895
8.6
<1
1.0
0.005
0.834
0.313
1,29
1,29
1,29
11.4
<1
<1
0.002
0.725
0.357
1,81
1,80
1,71
4.9
<1
<1
0.030
0.857
0.493
Region 10 (two)
Constraint
Congruency
Constraint · congruency
1,54
1,54
1,54
2309
2742
2495
6.2
<1
2.6
0.016
0.347
0.112
1,29
1,29
1,29
2.3
<1
2.3
0.142
0.414
0.141
1,50
1,69
1,72
1.7
<1
1.2
0.202
0.535
0.273
Region 11 (jobs to pay the bills)
Constraint
1,54
Congruency
1,54
Constraint · congruency
1,54
2883
2234
2183
1.4
1.1
1.4
0.236
0.295
0.249
1,29
1,29
1,29
<1
1.8
<1
0.579
0.196
0.545
1,42
1,82
1,45
<1
<1
<1
0.614
0.411
0.590
Region 12 (3rd subject)
Constraint
Congruency
Constraint · congruency
1,54
1,54
1,54
6854
9304
8821
<1
4.2
<1
0.820
0.046
0.970
1,29
1,29
1,29
<1
3.1
<1
0.754
0.087
0.762
1,83
1,68
1,56
<1
1.8
<1
0.853
0.185
0.971
Region 13 (never)
Constraint
Congruency
Constraint · congruency
1,54
1,54
1,54
3266
3381
3623
5.9
2.2
2.7
0.019
0.140
0.109
1,29
1,29
1,29
4.3
1.6
3.4
0.048
0.216
0.075
1,67
1,67
1,80
2.5
<1
1.5
0.121
0.337
0.226
Region 14 (got)
Constraint
Congruency
Constraint · congruency
1,54
1,54
1,54
1706
2514
1958
<1
3.6
<1
0.517
0.062
0.698
1,29
1,29
1,29
<1
3.2
<1
0.588
0.083
0.817
1,67
1,73
1,50
<1
1.7
<1
0.676
0.195
0.842
Region 15 (to see her)
Constraint
Congruency
Constraint · congruency
1,54
1,54
1,54
1339
1089
1513
<1
1.9
<1
0.920
0.178
0.983
1,29
1,29
1,29
<1
1.0
<1
0.707
0.329
0.897
1,61
1,59
1,57
0.0
0.6
0.0
0.922
0.425
0.983
394
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
Principle C conditions are presented in Fig. 1 and the
results from the no-constraint and name conditions in
Fig. 2.
Principle C and no-constraint conditions. There were no
reliable main effects or interactions in any region in the
first clause.
In the main ANOVA at the critical second subject in
region 8 there was a main effect of constraint due to
shorter mean reading times in the Principle C conditions
than in the no-constraint conditions (mean for the noconstraint conditions = 425.5 ms, mean for the Principle
C conditions = 377.3 ms, 95% CI = 30.3 ms). This effect
can be attributed to the differences in the content of the
preceding region between the two pairs of conditions.
Most importantly, there also was a significant constraint · congruency interaction at the second subject
in the participants analysis. Pairwise comparisons within
each level of the constraint factor showed no effect of
gender congruency for the Principle C conditions (gender-match—383.4 ms,
gender-mismatch—371.1 ms,
95% CI = 23.8 ms) and a marginally significant effect
of gender congruency for the no-constraint conditions
(gender-match—408.4 ms, gender-mismatch—442.6 ms,
95% CI = 40.3 ms) due to a slowdown in average reading times in the gender-mismatch condition (i.e., a mismatch effect). The main effect of constraint persisted to
the region immediately following the second subject
(mean for the no-constraint conditions = 390.7 ms,
mean for the Principle C conditions = 361.7 ms, 95%
CI = 18.3 ms). However, neither the main effect of congruency nor the constraint · congruency interaction
were significant and no significant differences were
observed in the pairwise comparisons within each level
of the factor constraint (Principle C conditions: gender-match—366.1 ms,
gender-mismatch—357.3 ms,
95% CI = 16.1 ms; no-constraint conditions: gendermatch—388.0 ms, gender-mismatch—393.5 ms, 95%
CI = 26.1 ms). The main effect of constraint was still significant in the second post-subject region in the participants analysis, although not in the items analysis
(mean for the no-constraint conditions = 351.4 ms,
mean for the Principle C conditions = 369.0 ms, 95%
CI = 13.0 ms).
Other reliable effects included a main effect of congruency at the third subject noun phrase in region 12
that was significant in the participants analysis and marginally significant in the items analysis, due to longer
reading times for the third subject in the match conditions
than in the mismatch conditions (mean for the gendermatch conditions = 450.2 ms, mean for the gender-mismatch conditions = 426.0 ms, 95% CI = 26.1 ms). Recall
that the congruency factor manipulated the (mis)match
in gender between the pronoun and the 2nd subject,
rather than the 3rd subject. There was also a main effect
of the constraint factor at the following word (mean for
the no-constraint conditions = 351.8 ms, mean for the
Principle C conditions = 371.9 ms, 95% CI = 15.4 ms)
due to slower reading times in the Principle C conditions. Pairwise comparisons in the same region showed
a significant effect of congruency for the Principle C
pair (gender-match—362.7 ms, gender-mismatch—
340.9 ms, 95% CI = 18.6 ms) due to longer reading
times in the match condition, but no corresponding
effect of congruency in the no-constraint pair (gendermatch—371 5 ms, gender-mismatch—372.3 ms, 95%
CI = 26.4 ms). No other differences were statistically
reliable. The effects in regions 12 and 13 do not receive
a straightforward explanation, e.g. they cannot be
explained by the parser’s anticipation of an antecedent
noun phrase in conditions where the anaphoric dependency has not yet been resolved (i.e. both Principle C
conditions and the gender-mismatch condition of the
no-constraint pair).
Name condition. In the first clause there was a significant difference in reading times between the name and
the mismatch condition of the no-constraint pair at the
subject noun phrase in region 4 (mean for the name condition = 377.4 ms, mean for the no-constraint mismatch
condition = 351.0 ms, 95% CI = 19.5 ms) and at the following region (mean for the name condition = 356.7 ms,
mean for the no-constraint mismatch condition = 331.9 ms, 95% CI = 16.7 ms), due to longer reading times in the name condition. These effects are
expected in light of the length differences in the first subject noun phrase, which was a personal pronoun in the
no-constraint conditions and a proper name in the name
condition.
Reading times at the critical second subject noun
phrase were significantly longer in the mismatch condition of the no-constraint pair than in the name condition (mean for the name condition = 393.9 ms, mean
for the no-constraint mismatch condition = 442.6 ms,
95% CI = 41.0 ms). However, reading times at the
same region did not differ between the name and the
match condition (mean for the name condition = 393.9 ms, mean for the no-constraint match condition = 408.4 ms, 95% CI = 22.6 ms). The fact that
reading times at the second subject were longer in the
mismatch condition than in the other two conditions
is unlikely to be a spill-over from an earlier region,
in light of the almost identical reading times across
all three conditions at the immediately preceding region
(i.e. region 6). There were additional reading-time differences between the name condition and either of
the no-constraint conditions at later regions. These differences are expected in light of the parser’s attempt to
resolve coreference at the second subject noun phrase
in the no-constraint conditions but not in the name
condition and are not reported here as they are not
germane to the goals of this study.
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
Discussion
The off-line ratings showed that participants judged
the second subject noun phrase to be an acceptable antecedent for the preceding pronoun in the no-constraint
condition, but not in the Principle C condition, confirming that the contrast between the conditions is real, as
predicted by the Principle C constraint. Notwithstanding the fact that the effect remained marginally significant, our results from the self-paced reading task
replicated Van Gompel and Liversedge’s (2003) gender-mismatch effect in the no-constraint conditions:
there was a slowdown in the second subject position if
that noun phrase mismatched in gender with the preceding pronoun. In contrast, there was no effect of gender
congruency at the corresponding position in the Principle C conditions, or in any other region in the second
clause. The absence of a mismatch effect in the Principle
C condition suggests that in this condition the parser
never considered the subject of the second clause as a
potential antecedent, thus making the gender-congruency
between that subject and the preceding pronoun irrelevant. We can also rule out the possibility that the gender-mismatch effect in the no-constraint condition was
due to the cost of adding a new discourse referent, since
reading times at the second subject in the name condition patterned with the gender-match condition rather
than with the gender-mismatch condition. The name
condition was no harder to process than the gender
match condition in any of the regions, so the difference
between the gender match and mismatch conditions cannot be due to the introduction of a new discourse referent in the mismatch condition.
The presence of the gender mismatch effect in the noconstraint conditions indicates that an anaphoric dependency is formed at the second subject noun phrase before
gender information is taken into account, but it does not
show exactly when the dependency is formed. This could
occur immediately following the identification of the category of the subject based on bottom-up information, as
in Van Gompel & Liversedge’s account. Alternatively,
dependency formation could occur before any bottomup information about the second subject noun phrase
appears, due to the fact that a sentence-initial adverbial
clause reliably predicts a higher clause that contains a
subject noun phrase. Both of these accounts may be
viewed as evidence for ‘active’ search for the antecedent
for a pronoun, but uncertainty about the role of prediction leads to a problem in interpreting the Experiment 1
results. The argument for the early on-line effects of Principle C involves the absence of a mismatch effect, and
therefore relies on the assumption that there is no other
difference between the conditions that could have led to
the absence of a mismatch effect in the Principle C conditions. However, there is an inherent asymmetry between
the Principle C and the no-constraint conditions related
395
to the parser’s ability to predict the second subject position. Recall that the sentences used in the experiment
were 3-clause structures, schematized in (14).
(14)
Schematic representation of conditions from
Experiment 1.
No-constraint:
[Because [while pronoun [2nd subject. . .. . ..]], 3rd
subject. . .]
Principle C:
[Because [pronoun. . ..[while 2nd subject. . ...]], 3rd
subject. . .]
In the no-constraint conditions the parser could reliably predict the second subject position immediately
after encountering the subordinator while in the first
clause. In contrast, the second subject position cannot
be reliably predicted during the first clause in the Principle C conditions. Furthermore, in these conditions the
third subject position, which corresponds to the main
clause subject, can be reliably predicted during the first
clause. Therefore, the reason for the absence of a gender
mismatch effect in the Principle C conditions could be
due to the unpredictability of the second subject position, rather than to the effects of the grammatical constraint. Furthermore, the lack of expectation for the
second subject position could strongly reduce any effect
of the gender of the noun phrase in that position.
To separate the effects of grammatical constraints
and predictability, Experiments 2 and 3 tested the effects
of the Principle C constraint in configurations where the
critical noun phrase position is equally predictable
across all conditions.
Experiment 2
The goal of Experiment 2 was to test the effect of
Principle C on active search for antecedents, while
avoiding confounds due to differences in the predictability of the critical subject noun phrase. In Experiment 2
both the Principle C and the no-constraint conditions
were identical with respect to the predictability of the
second clause at all times. Thus, a contrast between
the results for the no-constraint and Principle C conditions in this study would provide stronger evidence for
the impact of grammatical constraints on the construction of anaphoric dependencies.
Participants
Participants were 60 native speakers of English from
the University of Maryland undergraduate population
with normal or corrected-to-normal vision and no
history of language disorders.
396
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
Materials and design
The materials for the on-line task are described first,
since the materials for the rating study were derived
from them. The experiment had four conditions in a
2 · 2 design with the within-subjects factors constraint
(Principle C vs. no-constraint) and gender congruency
(match vs. mismatch between the pronoun and the subject of the second clause). A full set of items is shown in
Table 3.
All target sentences started with a clause with an
impersonal subject and a non-agentive predicate such
as it seemed or it was surprising. We reasoned that a sentence fragment like It seemed worrisome to . . . would create a strong expectation for a complement clause and
hence for an additional subject position. Consequently,
in all conditions the second subject position could be
anticipated when the cataphoric pronoun was first
encountered, thereby eliminating the potential confound
that made it more difficult to interpret the distribution of
gender mismatch effects in Experiment 1. In the Principle C conditions the cataphoric pronoun was the complement of the dative preposition to or for and referred
to the experiencer of the main clause predicate. In this
configuration coreference between the pronoun and the
subject of an embedded clause is unacceptable. In the
no-constraint conditions the pronoun was embedded
as the possessor inside the complement of to or for. In
this configuration the pronoun may corefer with the subject of an embedded clause. Note that this contrast in
acceptability follows from the syntactic formulation of
Principle C only if prepositions are assumed to be irrelevant for calculation of c-command relations. This
assumption is consistent with results from other diagnostics of c-command, such as negative polarity item
licensing and scope assignment, and the reader is
referred to Reinhart (1983) and Brody (1994) for further
discussion. As in Experiment 1, the materials were
designed such that a cataphoric pronoun could always
find a grammatically accessible intra-sentential antecedent, even in the Principle C conditions. This was
achieved by adding a but-clause to the end of the sentence, the subject of which was a licit antecedent for
the pronoun. The gender of the third subject was chosen
such that there was a unique antecedent for the pronoun
in each condition. The stimuli were designed such that
plausibility considerations should not favor coreference
between the pronoun and the embedded subject more
strongly in either the Principle C condition or the noconstraint condition (irrespective of grammatical acceptability). For example, in the set of materials in Table 3
both John and his family could plausibly be worried
about something negative involving John such as gaining weight.
Thirty two sets of target items were distributed
among four presentation lists in a Latin Square design.
Each list contained 32 experimental sentences (8 per
condition) and 90 filler sentences of varying length and
complexity, thereby maintaining target-to-filler ratio of
approximately 1:3. Filler sentences were identical across
all four lists, and were designed with the same considerations used in Experiment 1. Each participant was randomly assigned to one of the lists, and the order of the
stimuli within the presentation list was randomized for
each participant. The full list of materials is given in
Appendix B.
Acceptability rating task (off-line)
An off-line rating task was used to confirm that, provided matching gender/number values, participants
accept coreference between the cataphoric pronoun
and the second subject in the no-constraint environment,
but reject it in the Principle C environments. The acceptability rating task for Experiment 2 was run together
with the corresponding task for Experiment 1, and so
was completed by the same 40 participants described
there, using the same design and procedure. Participants
were asked to rate the acceptability of coreference on a
scale from 1 (impossible) to 5 (absolutely natural). The
test items were 12 pairs of gender-match conditions chosen from the materials for the online task, distributed
Table 3
Sample set of experimental items for on-line Experiment 2
Principle C/match
It seemed worrisome to himi that John was gaining so much weight, but Matti didn’t have the nerve
to comment on it
Principle C/mismatch
It seemed worrisome to himi that Ruth was gaining so much weight, but Matti didn’t have the nerve
to comment on it
No constraint/match
It seemed worrisome to hisi family that Johni was gaining so much weight, but Ruth thought it was
just a result of aging
No constraint/mismatch
It seemed worrisome to hisi family that Ruth was gaining so much weight, but Matti thought it was
just a result of aging
The underlined name indicates the critical second subject noun phrase. Subscript indices indicate intended backward anaphoric
dependencies.
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
Results
across two presentation lists using a Latin Square design
and interspersed with 48 filler items (additional to the
materials from Experiment 1).
The mean rating score for the Principle C condition
(mean = 1.5, standard error = .12) was significantly
lower than in the no-constraint condition (mean = 4.2,
standard error = .13) both in the participants and items
analyses (two-tailed paired t-test, both ps < .001,
t1(39) = 12.9, t2(11) = 31.2). These results strongly suggest that participants recognize the constraint on coreference for the Principle C structures in (14).
Comprehension question accuracy
The mean question answering accuracy was 95.7%
and ranged between 94.4% and 97.3% for individual
conditions. There were no reliable differences in accuracy between individual conditions (Fs < 1).
Self-paced reading
The results from Experiment 2 are plotted in Fig. 3.
The results of the 2 · 2 ANOVA analysis are reported
in Table 4.
At the impersonal subject it in region 1 the general
ANOVA showed an interaction of constraint and congruency that was significant in the participants analysis
and marginally significant in the items analysis. However, this interaction was not supported in the pairwise
comparisons within each level of the constraint factor.
There were no other significant effects or interactions
anywhere in the first clause.
At the complementizer that in region 7 there was a
main effect of the constraint factor, due to longer reading times for the no-constraint conditions than for the
Principle C conditions (mean for the no-constraint conditions = 321.7 ms, mean for the Principle C conditions = 302.4 ms, 95% CI = 7.7 ms). This difference
was likely due to the lexical difference between conditions at the preceding regions. A main effect of the constraint factor in the same direction was also found at the
subsequent region, the critical subject of the second
clause (region 8: mean for the no-constraint conditions
= 328.5 ms, mean for the Principle C conditions =
317.7 ms, 95% CI = 10.3 ms). However, in region 8 the
constraint · congruency interaction did not reach significance, and planned pairwise comparisons revealed no
Procedure and analysis
The procedure was a self-paced word-by-word
moving window task, identical to Experiment 1. Each
sentence was followed by a yes-no comprehension
question.
Data from all 60 participants was included in the analysis, which followed the same steps as in Experiment 1.
Sentences for which the comprehension question was
answered incorrectly were excluded from the analysis,
and reading times that exceeded a threshold of 2.5
standard deviations above a participant’s mean reading
rate for each region were replaced by the threshold
value. This winsorizing procedure affected 2.0% of trials
(1.9–2.2% for individual conditions).
The regions used for the data analysis corresponded
to single words, except for regions corresponding to
the end of the clause, for which several words were averaged together due to variation in the clause length
between items (see the legend for Fig. 3 for region
encoding). Raw reading times for each region were
entered into a 2 · 2 repeated-measures ANOVA with
the factors constraint and gender congruency.
Principle C, match
360
397
subject
NP
Principle C, mismatch
Mean reading time (ms)
No constraint, match
No constraint, mismatch
340
320
300
280
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
It1 seemed2 worrisome3 to4 {him5/ his5 family6} that7 John/Ruth8 was9 gaining10 so11
(much weight)12, but13 Matt14 didn't15 have16 (the nerve to comment on it)17.
Fig. 3. Mean reading times in milliseconds from Experiment 2. The error bars represent one standard error above/below the
participant mean at each region. The arrow marks the critical second subject noun phrase.
398
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
Table 4
Results of 2 · 2 ANOVAs and Min F 0 values for Experiment 2
By participants
Min F 0
By items
df
F2
p
df
min F 0
p
0.626
0.564
0.571
1,31
1,31
1,31
<1
<1
<1
0.754
0.690
0.750
1,31
1,31
1,31
<1
<1
<1
0.792
0.743
0.781
5.4
<1
<1
0.024
0.821
0.743
1,31
1,31
1,31
3.2
<1
<1
0.082
0.754
0.701
1,31
1,31
1,31
2.0
<1
<1
0.165
0.855
0.803
1448.3
997.3
1401.9
16.3
<1
<1
<0.001
0.486
0.912
1,31
1,31
1,31
10.6
<1
<1
0.003
0.654
0.966
1,31
1,31
1,31
6.4
<1
<1
0.016
0.707
0.968
1,59
1,59
1,59
1589.5
1187.6
1624.0
4.7
3.2
<1
0.034
0.077
0.927
1,31
1,31
1,31
4.3
2.7
<1
0.048
0.112
0.895
1,31
1,31
1,31
2.2
1.5
<1
0.145
0.235
0.940
Region 9 (was)
Constraint
Congruency
Constraint · congruency
1,59
1,59
1,59
1925.5
2223.8
1681.0
<1
3.7
<1
0.880
0.059
0.544
1,31
1,31
1,31
<1
3.9
<1
0.708
0.058
0.406
1,31
1,31
1,31
<1
1.9
<1
0.889
0.179
0.625
Region 10 (gaining)
Constraint
Congruency
Constraint · congruency
1,59
1,59
1,59
1422.0
1055.7
1587.0
1.6
8.8
<1
0.209
0.004
0.601
1,31
1,31
1,31
1.3
5.0
<1
0.260
0.033
0.438
1,31
1,31
1,31
<1
3.2
<1
0.401
0.085
0.665
Region 11 (so)
Constraint
Congruency
constraint · congruency
1,59
1,59
1,59
1339.5
1909.9
2056.6
8.1
7.4
6.7
0.006
0.009
0.012
1,31
1,31
1,31
5.2
5.7
9.0
0.030
0.023
0.005
1,31
1,31
1,31
3.2
3.2
3.8
0.085
0.083
0.059
Region 12 (much weight)
Constraint
Congruency
Constraint · congruency
1,59
1,59
1,59
3863.8
5224.2
2737.8
4.0
1.9
1.1
0.049
0.175
0.297
1,31
1,31
1,31
3.1
5.0
1.9
0.090
0.033
0.183
1,31
1,31
1,31
1.7
1.4
<1
0.196
0.250
0.411
Region 13 (but)
Constraint
Congruency
Constraint · congruency
1,59
1,59
1,59
1464.2
2053.6
2752.1
1.6
<1
2.4
0.210
0.450
0.126
1,31
1,31
1,31
1.0
1.3
5.9
0.326
0.256
0.021
1,31
1,31
1,31
<1
<1
1.7
0.439
0.530
0.200
Region 14 (3rd subject)
Constraint
Congruency
Constraint · congruency
1,59
1,59
1,59
1106.5
1562.1
1231.9
<1
5.8
<1
0.480
0.019
0.383
1,31
1,31
1,31
<1
8.5
<1
0.578
0.007
0.367
1,31
1,31
1,31
<1
3.5
<1
0.662
0.073
0.530
Region 15 (didn’t)
Constraint
Congruency
Constraint · congruency
1,59
1,59
1,59
1399.0
1101.5
1840.9
2.5
8.8
<1
0.122
0.004
0.789
1,31
1,31
1,31
1.4
3.0
<1
0.240
0.091
0.764
1,31
1,31
1,31
<1
2.3
<1
0.349
0.143
0.842
Region 16 (have)
Constraint
Congruency
Constraint · congruency
1,59
1,59
1,59
1134.4
1309.7
1176.0
3.8
4.0
5.5
0.058
0.050
0.022
1,31
1,31
1,31
1.0
4.5
5.6
0.332
0.042
0.024
1,31
1,31
1,31
<1
2.1
2.8
0.387
0.156
0.105
df
MSE
F1
Region 4 (to)
Constraint
Congruency
constraint · congruency
1,59
1,59
1,59
1279.1
943.0
880.6
<1
<1
<1
Region 5 (him/his)
Constraint
Congruency
Constraint · congruency
1,59
1,59
1,59
955.4
944.8
820.4
Region 7 (that)
Constraint
Congruency
Constraint · congruency
1,59
1,59
1,59
Region 8 (2nd subject)
Constraint
Congruency
Constraint · congruency
p
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
effect of congruency within either level of the constraint
factor.
The ANOVAs at each of the three regions immediately following the subject (regions 9–11 in Fig. 1)
showed a significant or a marginally significant main
effect of congruency, due to slower reading times in mismatching conditions than matching conditions (region 9:
mean for the gender-match conditions = 321.6 ms, mean
for the gender-mismatch conditions = 333.8 ms, 95%
CI = 12.2 ms; region 10: mean for the gender-match
conditions = 324.3 ms, mean for the gender-mismatch
conditions = 337.2 ms, 95% CI = 8.4 ms; region 11:
mean for the gender-match conditions = 323.4 ms, mean
for the gender-mismatch conditions = 338.5 ms, 95%
CI = 11.3 ms). In region 11 there also was a significant
main effect of constraint (mean for the no-constraint
conditions = 337.8 ms, mean for the Principle C conditions = 324.1 ms, 95% CI = 11.7 ms) and a significant
constraint · congruency interaction due to longer reading times in the no-constraint mismatch condition (noconstraint
conditions:
gender-match—322.4 ms,
gender-mismatch—353.1 ms; Principle C conditions:
gender-match—324.4 ms, gender-mismatch—323.9 ms,
95% CI = 23.4 ms). Planned pairwise comparisons at
regions 9–11 showed a significant or marginally significant effect of gender congruency (i.e., gender-mismatch
effect) in the no-constraint conditions at all three regions
(region 9: mean for the gender-match condition = 320.2 ms, mean for the gender-mismatch condition = 336.4 ms, 95% CI = 16.8 ms; region 10: 325.8,
341.9 and 14.9 ms; region 11: 322.4, 353.1 and 19.2 ms,
respectively). The effect of gender congruency did not
reach significance in the Principle C conditions in any
region in the second clause, although there was a numerical trend towards slower reading times in the incongruent conditions.
In the third clause there was a significant main effect
of congruency in the overall ANOVA at the subject
noun phrase in region 14, due to longer reading times
in the mismatch conditions than in the match conditions
(mean for the gender-match conditions = 311.9 ms,
mean for the gender-mismatch conditions = 323.9 ms,
95% CI = 10.2 ms). Pairwise comparisons showed that
the effect of gender congruency was significant in the
Principle C conditions (region 14: mean for the gender-match condition = 311.9 ms, mean for the gendermismatch condition = 327.7 ms, 95% CI = 12.7 ms),
but not in the no-constraint conditions (region 14: mean
for the gender-match condition = 312.0 ms, mean for
the gender-mismatch condition = 320.2 ms, 95% CI =
14.6 ms). This was the first region where an accessible
antecedent for the pronoun could be found in all but
the no-constraint/match condition. This pattern of reading-time contrasts was unexpected, especially in light of
the persistency of the effect that continued to hold in
regions 15 and 16.
399
Discussion
The goal of Experiment 2 was to test the claim that
Principle C immediately restricts dependency formation
during online processing of backwards anaphora, using
materials that match the predictability of potential antecedent positions across conditions. As in Experiment 1,
this experiment elicited a gender mismatch effect in the
no-constraint conditions, and no corresponding effect
in the Principle C conditions. These findings support
the claim that the parser does not search for an antecedent for a pronoun in grammatically inaccessible positions, and do not straightforwardly lend themselves to
an explanation in terms of predictability, unlike Experiment 1.
However, it is clear that the mismatch effect in the noconstraint conditions in Experiment 2 was both delayed
and diminished in comparison to Experiment 1. The gender mismatch effect showed a significant or marginally
significant mismatch effect at the each of the three words
following the critical region, but the overall constraint · congruency interaction did not reach significance until the third region after the critical noun
phrase.
One possible reason for the smaller and delayed gender-mismatch effect in this experiment relative to Experiment 1 involves the structural necessity of the clause
that contains the potential antecedent noun phrase.
Experiment 1 introduced pronouns in adverbial clauses
introduced by subordinators like while, which are obligatorily followed by a main clause. The target items in
Experiment 2 introduced pronouns inside a main clause
with a beginning like It seemed clear. . ., which were
assumed to create a strong bias for a clausal continuation, based on intuition and on informal corpus searches
(Linguist’s Search Engine: Resnik & Elkiss, 2003). However, the lead-ins used in Experiment 2 are compatible
with a grammatical (although pragmatically unlikely)
parse as mono-clausal sentences with referential subject
pronouns. To test whether the weaker gender mismatch
effect in this experiment might have been due to a weaker expectation for a following clause, we conducted an
offline sentence completion task to determine the
strength of the prediction for a complement clause.
Thirteen undergraduate students from the University
of Maryland completed a pencil-and-paper completion
task for monetary compensation. The questionnaire
contained sentence-initial fragments such as (15), which
participants were asked to complete in a natural fashion.
The last word of the fragment was a possessive pronoun,
e.g. his. The pronoun was included to assess what proportion of the completions that contained an embedded
clause also contained a referent for the pronoun as part
of the embedded clause. All target fragments contained
the masculine 3rd person singular pronoun his or plural
their. The feminine pronoun her was not used in the
400
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
target sentences due to the homophony of the accusative
and possessive feminine forms. The questionnaire consisted of 12 target fragments, interspersed with 60 fragments of other types.
(15) It seemed clear to his. . ..
The results from the completion task suggest that a
sentence fragment such as (15) created a strong but
not absolute expectation for an upcoming embedded
clause. Participants provided a clausal continuation in
69% (99/143) of cases (71% (51/72) in the ‘his’ condition,
67% (48/71) in the ‘their’ condition), but in the remaining 29% (44/143) of cases the completed sentence was a
simple monoclausal sentence. Such monoclausal completions are grammatical, since the sentence-initial it
can be understood as a 3rd person neuter pronoun with
an unspecified referent, but they are pragmatically infelicitous in the absence of a supporting context. The complement clause contained a potential antecedent for the
cataphoric pronoun his or their in 70% (69/99) of the
completions that included a second clause, suggesting
that when the embedded clause was expected, a referent
for the pronoun was also often expected. In 87% (60/69)
of the completions that provided a potential antecedent,
the potential referent appeared in subject position.
The results from the fragment completion questionnaire lend support to the notion that speakers normally
expect a clausal continuation following a lead-in like It
seemed clear to his . . ., and also suggest that in most
cases speakers also anticipate an antecedent for the pronoun. However, the 30% of completions that did not
contain an embedded clause still represent a significant
difference from the materials in Experiment 1, where a
completion to a fragment like While he . . . would necessarily include a second clause. Thus, it is possible that
the difference in strength of prediction represented by
this 30% contributed to the weakened gender mismatch
effect in Experiment 2.
Therefore, Experiment 2 provided further evidence
that active search for antecedents for pronouns is
restricted by grammatical constraints on coreference,
while leaving open certain questions about the strength
and immediacy of gender-mismatch effects and what
they show about the time-course of anaphoric dependency formation. In particular, the results of Experiment
2 show that the distribution of gender-mismatch effects
cannot be explained entirely in terms of predictability,
but do not indicate whether the predictability of a potential antecedent position is a prerequisite for eliciting a
gender-mismatch effect. The results of Experiments 1
and 2 are consistent with a view in which active dependency formation mechanisms for cataphora are activated
only at the time when the pronoun is encountered.
Under this view, which we refer to as instantaneous resolution, the parser actively constructs anaphoric depen-
dencies only in positions that can be reliably predicted
at the time when the pronoun is encountered. Instantaneous resolution predicts that a dependency between
the anaphoric pronoun (his) and the second subject position (John) would be actively constructed in a structure
like ‘While his mother slept, John sang’, in which the
second subject position can be anticipated even before
the pronoun is encountered, due to the early appearance
of the subordinator while. On the other hand, instantaneous resolution predicts that the parser would not
attempt a similar dependency in ‘His mother slept, while
John sang’, as the second subject position cannot be reliably predicted until after the pronoun is encountered.
An alternative to the instantaneous resolution view is
that the parser’s search for candidate antecedent positions remains active beyond the point when the pronoun
is encountered and takes account of any new piece of
structure predicted by bottom-up information following
the pronoun. This view, which we refer to as continuous
resolution, predicts that the parser would actively consider a dependency between a cataphoric pronoun and
the second subject even in ‘His mother slept, while John
sang’, despite the fact that a reliable prediction for the
second subject position becomes available only with
the subordinator while and after the anaphoric pronoun
is encountered. This latter view is a closer parallel to the
properties of active search for gaps in processing fillergap dependencies, where active gap creation effects have
been found in positions that could not have been anticipated at the point when the filler was first encountered
(e.g., Aoshima et al., 2004; Frazier & Clifton, 1989).
Importantly, whereas the two views disagree on the
specific configuration that must obtain to initiate an
active search, they both agree that a cataphoric dependency can be formed based on a prediction for an antecedent position and before the antecedent becomes
available bottom-up. We assume that speakers can
encode the fact that a pronoun is linked to another syntactic position that serves as its antecedent, independent
of encoding what the antecedent refers to. We refer to
this process as ‘linking’, and suggest that it may occur
as soon as an antecedent position is predicted. We
assume that a separate process of ‘valuation’ occurs
when the properties of the antecedent are processed bottom-up, and the pronoun automatically inherits the referential properties of the antecedent.
Experiment 3 attempts to adjudicate between the
instantaneous resolution and the continuous resolution
proposals by testing a configuration in which no syntactic prediction for a potential antecedent position is available at the time when the cataphoric pronoun is
encountered. If a gender mismatch effect is found in such
a configuration, it would be evidence that the search
mechanism does not search for candidate antecedents
just once, but repeatedly, as more information comes
online. Therefore, the results of this third experiment
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
will also speak further to the extent of the parallelism
between active search mechanisms in the two construction types.
Experiment 3
The goals of Experiment 3 were twofold. First, to test
the instantaneous vs. continuous resolution accounts,
i.e., whether active search for pronoun antecedents
occurs in structural positions that could not be reliably
predicted at the point when the pronoun is first encountered. Second, as in Experiments 1 and 2 we sought to
test whether active search for pronoun antecedents is
restricted by grammatical constraints, specifically Principle C.
Participants
Participants were 60 native speakers of English from
the University of Maryland undergraduate population
with normal or corrected-to-normal vision and no history of language disorders. They gave informed consent
and were financially compensated for their participation.
Materials and design
The target items for Experiment 3 consisted of 24 sets
of 4 items organized in a 2 · 2 design with the factors
constraint (no-constraint vs. Principle C) and congruency
(match vs. mismatch) as within-subjects factors. One full
set of items is shown in Table 5 and a complete set of
materials can be found in Appendix C.
All conditions began with a main clause that was followed by an adverbial clause headed by a subordinator
such as while. The pronoun appeared as the subject of
the main clause in the Principle C conditions, or as the
possessor of the main clause subject in the no-constraint
conditions. The critical potential antecedent position
401
was the subject noun phrase of the embedded adverbial
clause, which could not be reliably predicted at the point
when the pronoun was first encountered. The first possible cue for such a position appeared later in the sentence
when the subordinator while was encountered. Importantly, the subordinator while did not immediately precede the critical head noun of the second subject noun
phrase, but instead was separated from the head noun
by the determiner the and two adjectives. This ensured
that the parser had ample opportunity to process the
information associated with the subordinator before
encountering the critical noun, i.e., (i) that there is a
new clause and, hence, an upcoming subject position,
and (ii) that in the Principle C condition this subject
position is c-commanded by the cataphoric pronoun
and thus excluded as a potential antecedent due to Principle C.
Gender congruent and incongruent sentences were
formed by manipulating the gender of the sentence-initial pronoun such that it either matched or mismatched
the head noun of the second subject noun phrase. The
nouns used were either lexically (e.g. king, queen) or conventionally (e.g. quarterback) strongly gender-specific.
To make it impossible to reliably anticipate the gender
of the noun, the adjectives preceding the noun did not
include gender-biased adjectives (e.g., pregnant or
handsome).
As in the previous experiments, to ensure the possibility of intra-sentential coreference, the sentences
included an additional clause that contained a licit antecedent for the cataphoric pronoun whenever no suitable antecedent was available in the second clause. The twenty-four
sets of experimental stimuli were distributed across four
lists in a Latin Square design. Each list contained 24
experimental sentences and 72 filler sentences. Filler sentences were identical across all four lists. Each participant was randomly assigned to one of the lists, and
the order of the stimuli within the presentation list was
randomized for each participant.
Table 5
Sample set of experimental items for on-line Experiment 3
Principle C/match
Hei chatted amiably with some fans while the talented, young quarterback signed autographs for
the kids, but Stevei wished the children’s charity event would end soon so he could go home
Principle C/mismatch
Shei chatted amiably with some fans while the talented, young quarterback signed autographs for
the kids, but Caroli wished the children’s charity event would end soon so she could go home
No constraint/match
Hisi managers chatted amiably with some fans while the talented, young quarterbacki signed
autographs for the kids, but Carol wished the children’s charity event would end soon so
she could go home
No constraint/mismatch
Heri managers chatted amiably with some fans while the talented, young quarterback signed autographs
for the kids, but Caroli wished the children’s charity event would end soon so she could go home
The underlined noun indicates the head of the critical second subject noun phrase. Subscript indices indicate intended backward
anaphoric dependencies.
402
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
Acceptability rating task (off-line)
Procedure and analysis
As in Experiments 1 and 2, an off-line acceptability
rating task was used to confirm that the configurations
in the no-constraint and Principle C conditions effectively varied the acceptability of coreference between the
pronoun and the subject of the second clause. Participants were the same 60 native speakers of English who
participated in the on-line portion of Experiment 3. To
avoid bias from the acceptability rating task on the
on-line results, participants completed the rating questionnaire after the on-line experiment. The task was
identical to that used in Experiments 1 and 2, i.e., participants rated the acceptability of coreference on a scale
from 1 (impossible) to 5 (absolutely natural). Twentyfour sets of stimuli were constructed based on 24 target
stimulus sets from on-line Experiment 3. The Principle C
and the no-constraint conditions were derived from the
gender-match conditions of the on-line study by removing the sentence-final but-clause. The forwards anaphora
condition was created by exchanging the pronoun and
the second subject noun phrase in the Principle C condition. This condition was included to address the possible
concern that lower acceptability ratings in the Principle
C condition might reflect the implausibility of having the
main and embedded clause events be simultaneously
performed by the same agent. The but-condition was
identical to the Principle C/match condition in the online study and was included to test the effectiveness of
our effort to provide a licit intra-sentential antecedent
for the pronoun through the subject of an additional
but-clause.
Four experimental lists were constructed using a
Latin Square design, such that each participant saw only
one condition from each set. In addition to the 24 experimental sentences, each questionnaire also contained 36
filler sentences that were identical across the four lists.
The mean coreference rating scores from Experiment 3
are summarized in Table 6.
The Principle C condition received a mean rating
score of 1.7 that was significantly lower than the rating
score in the other three conditions (2-tailed paired t-test,
all ps < .01). The coreference rating score in the no-constraint and the but-conditions was significantly lower
than in the forwards anaphora condition, but this is
expected given that forwards anaphora is the preferred
way of expressing coreference in these contexts.
The procedure was a self-paced word-by-word moving window task, identical to Experiments 1 and 2. Each
sentence was followed by a yes-no comprehension
question.
The data from 2 of the 60 participants could not be
included due to technical problems. To balance the number of participants across each list, one participant was
excluded from each list, such that analysis was performed on 56 participants, distributed equally among
lists. As in Experiment 1, trials on which the comprehension question was answered incorrectly were excluded,
and reading-times that exceeded a threshold of 2.5 standard deviations above the mean reading rate for each
region and participant were replaced by that threshold
value. This procedure affected 2.2% of trials (2.1–2.3%
for individual conditions). The data were entered into
a 2 · 2 repeated-measures ANOVA.
Table 6
Mean rating scores from Experiment 3
Condition
Principle C
No-constraint
Forwards anaphora
But-condition
Mean rating (Standard error)
1.7
3.4
4.3
3.9
(.09)
(.13)
(.08)
(.09)
Results
The mean accuracy on comprehension question was
92.6% (91.6-93.8% for individual conditions). There
were no reliable differences in accuracy between individual conditions (Fs < 1).
The results from all conditions in Experiment 3 are
shown in Fig. 4. The results of the 2 · 2 ANOVAs are
summarized in Table 7.
In the first clause, there was a significant main effect
of the constraint factor at the verb-adverb sequence in
regions 3 and 4 due to longer reading times in the noconstraint conditions than in the Principle C conditions
(region 3: mean for the no-constraint conditions = 406.3 ms, mean for the Principle C conditions = 338.6 ms, 95% CI = 18.3 ms; region 4: 381.2,
345.1 and 14.6 ms respectively). In region 5 there was
a significant main effect both of constraint (mean for
the no-constraint conditions = 358.0 ms, mean for the
Principle C conditions = 338.7 ms, 95% CI = 15.1 ms)
and of congruency (mean for the gender-match conditions = 340.9 ms, mean for the gender-mismatch conditions = 355.8 ms, 95% CI = 8.8 ms), but no significant
interaction of the factors. These effects were partly due
to lexical differences between conditions at the preceding
regions and partly unexpected. At the conjunction while
in region 6 there was a significant main effect of constraint (mean for the no-constraint conditions =
362.3 ms, mean for the Principle C conditions =
337.8 ms, 95% CI = 12.8 ms) and a significant constraint · congruency interaction. Pairwise comparisons
revealed that the interaction was caused by an unexpected effect of congruency in the Principle C conditions that
contained identical lexical material until that point
(region 6: mean for the gender-match condition =
328.3 ms, mean for the gender-mismatch condition =
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
Mean reading time (ms)
460
subject
noun
403
Principle C, match
Principle C, mismatch
No constraint, match
420
No constraint, mismatch
380
340
300
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19
He (she)1 / [Principle C] or His (her)1 / managers2 / [no-constraint]
/ chatted3 / amiably4 / with some fans5 / while6 / the7 / talented,8 / young9 / quarterback10
/ signed11 / autographs12 /for13 /the kids,14 / but15 / Steve (Carol)16 / wished17 / the18 / children's
charity event would end soon so he could go home19.
Fig. 4. Mean reading times for the Principle C and no-constraint conditions from Experiment 3. The arrow marks the position of the
critical gender-marked head-noun. The error bars represent one standard error above/below the participant mean at each region. The
arrow marks the critical second subject noun phrase.
347.4 ms, 95% CI = 17.0 ms) that was not observed in
the no-constraint conditions (mean for the gendermatch condition = 369.6 ms, mean for the gender-mismatch condition = 354.9 ms, 95% CI = 17.2 ms). The
effect of constraint was marginally significant at the
determiner in region 7 (mean for the no-constraint
conditions = 326.0 ms, mean for the Principle C conditions = 316.9 ms, 95% CI = 9.6 ms), and was significant
at the first adjective in region 8 (mean for the no-constraint conditions = 354.1 ms, mean for the Principle C
conditions = 331.6 ms, 95% CI = 16.6 ms). Importantly,
there were no effects of congruency in the two regions
preceding the critical noun and similarly, no effects of
the constraint factor in the region immediately preceding
the critical noun.
At the critical noun in region 10 there was a main
effect of congruency (mean for the gender-match conditions = 367.1 ms, mean for the gender-mismatch conditions = 389.4 ms, 95% CI = 14.1 ms) and a significant
constraint · congruency interaction. Separate pairwise
comparisons of the Principle C and no-constraint conditions revealed a strong effect of congruency in the noconstraint pair in the predicted direction (region 10:
mean for the gender-match condition = 364.6, mean
for the gender-mismatch condition = 402.5, 95%
CI = 18.1 ms). No corresponding effect in the Principle
C pair (region 10: mean for the gender-match condition = 369.6 ms, mean for the gender-mismatch condition = 376.4 ms, 95% CI = 19.9 ms).
The constraint · congruency interaction was also significant at the word following the subject noun in region
11. Once again, pairwise comparisons within each level
of the constraint factor showed that the interaction
was due to the presence of a significant effect of congru-
ency in the no-constraint conditions (mean for the
gender-match condition = 384.4, mean for the gendermismatch condition = 417.0, 95% CI = 24.9 ms), and
no corresponding effect in the Principle C conditions
(mean for the gender-match condition = 395.6 ms, mean
for the gender-mismatch condition = 391.3 ms, 95%
CI = 22.4 ms). There were no other significant effects
in the remainder of the second clause or anywhere in
the final clause.
Discussion
The main finding of Experiment 3 was the presence of
a gender mismatch effect in the no-constraint condition
at the critical gender-marked head noun. The immediacy
and robustness of this effect in Experiment 3 is important in light of the fact that the corresponding effect
was marginally significant in Experiment 1 and delayed
in Experiment 2. The gender mismatch effect was present
despite the fact that there was no independent prediction
for an antecedent position at the time when the cataphoric pronoun was encountered. This suggests that
the parser’s active search for pronoun antecedents is
not a strictly instantaneous process that occurs when
the pronoun is first encountered, but is rather a continuous process that applies whenever it becomes possible
to anticipate a new potential antecedent position. In this
respect, therefore, active search for pronoun antecedents
resembles active search for gaps in filler-gap
dependencies.
In contrast, the gender mismatch effect was not found
at the critical second subject or in any other region in the
second clause in the Principle C conditions. These results
corroborate the conclusion drawn from Experiments 1
404
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
Table 7
Results of 2 · 2 ANOVAs and min F 0 values for Experiment 3
By participants
Min F 0
By items
df
MSE
F1
p
df
F2
p
df
min F 0
p
Region 6 (while)
Constraint
Congruency
Constraint times congruency
1,55
1,55
1,55
2283.0
2230.7
1843.1
13.7
<1
7.6
0.001
0.825
0.008
1,24
1,24
1,24
14.9
<1
11.2
0.001
0.832
0.003
1,69
1,65
1,75
7.1
<1
4.5
0.009
0.878
0.037
Region 7 (the)
Constraint
Congruency
Constraint times congruency
1,55
1,55
1,55
1293.0
1205.2
1500.1
3.6
<1
<1
0.062
0.936
0.652
1,24
1,24
1,24
4.2
<1
<1
0.051
0.851
0.896
1,71
1,71
1,28
2.0
<1
<1
0.167
0.941
0.900
Region 8 (talented)
Constraint
Congruency
Constraint times congruency
1,55
1,55
1,55
3848.1
2863.8
3604.4
6.2
1.3
1.4
0.016
0.260
0.250
1,24
1,24
1,24
11.4
1.5
1.1
0.003
0.229
0.316
1,78
1,71
1,60
4.0
<1
<1
0.049
0.405
0.445
Region 9 (young)
Constraint
Congruency
Constraint times congruency
1,55
1,55
1,55
3726.0
1963.0
3876.2
1.0
<1
<1
0.326
0.779
0.655
1,24
1,24
1,24
1.2
<1
<1
0.280
0.928
0.576
1,72
1,29
1,77
<1
<1
<1
0.463
0.931
0.725
Region 10 (quarterback)
Constraint
Congruency
Constraint times congruency
1,55
1,55
1,55
2772.2
3077.2
1949.5
2.5
8.7
8.1
0.118
0.005
0.006
1,24
1,24
1,24
3.0
4.9
6.2
0.097
0.037
0.021
1,71
1,52
1,60
1.4
3.1
3.5
0.246
0.082
0.066
Region 11 (signed)
Constraint
Congruency
Constraint times congruency
1,55
1,55
1,55
5072.1
3125.3
4710.2
<1
2.9
4.2
0.468
0.095
0.045
1,24
1,24
1,24
<1
2.8
5.7
0.431
0.110
0.026
1,71
1,66
1,74
<1
1.4
2.4
0.591
0.239
0.124
Region 12 (authographs)
Constraint
Congruency
Constraint times congruency
1,55
1,55
1,55
2112.0
1405.5
2298.9
<1
3.9
2.0
0.980
0.052
0.161
1,24
1,24
1,24
<1
<1
1.3
0.935
0.408
0.258
1,65
1,33
1,56
<1
<1
<1
0.981
0.444
0.373
Region 13
Constraint
Congruency
Constraint times congruency
1,55
1,55
1,55
1642.1
1175.3
1593.0
4.4
<1
<1
0.041
0.866
0.886
1,24
1,24
1,24
2.9
<1
<1
0.103
0.845
0.606
1,56
1,74
1,63
1.7
<1
<1
0.193
0.898
0.890
Region 14
Constraint
Congruency
Constraint times congruency
1,55
1,55
1,55
2114.9
2606.5
3352.8
1.3
4.3
<1
0.259
0.042
0.562
1,24
1,24
1,24
<1
1.8
<1
0.407
0.192
0.871
1,51
1,45
1,28
<1
1.3
<1
0.499
0.263
0.875
Region 15 (but)
Constraint
Congruency
Constraint times congruency
1,55
1,55
1,55
1847.7
1536.3
1672.7
<1
<1
<1
0.748
0.740
0.460
1,24
1,24
1,24
<1
<1
<1
0.617
0.742
0.794
1,79
1,67
1,30
<1
<1
<1
0.786
0.814
0.805
Region 16 (3rd subject)
Constraint
Congruency
Constraint times congruency
1,55
1,55
1,55
1530.2
1227.0
1604.4
1.5
<1
<1
0.224
0.987
0.398
1,24
1,24
1,24
1.9
<1
<1
0.183
0.865
0.493
1,72
1,56
1,56
<1
<1
<1
0.363
0.987
0.592
Region 17
Constraint
Congruency
Constraint times congruency
1,55
1,55
1,55
1992.0
1251.7
1799.6
1.6
0.1
0.2
0.205
0.704
0.664
1,24
1,24
1,24
1.8
0.2
0.0
0.190
0.661
0.911
1,69
1,74
1,27
0.9
0.1
0.0
0.355
0.773
0.914
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
and 2 that the initial set of antecedents for a cataphoric
pronoun does not contain noun phrase candidates that
violate Principle C. Note that this does not necessarily
exclude the possibility that the ‘illicit’ noun phrases
might be considered at some later point, before they
are ultimately rejected, although we find no evidence
of such ‘late’ effects (see Sturt, 2003 for evidence of late
effects involving forwards anaphora).
General discussion
Our primary concern in this study was to investigate
the extent to which mechanisms for processing backwards anaphoric dependencies match the mechanisms
that have been previously reported for the processing
of filler-gap dependencies, with a focus on mechanisms
of active search for pronoun antecedents and on the
impact of grammatical constraints on coreference relations. Results from our three off-line and three on-line
experiments together yield a fairly consistent picture.
The three off-line rating studies all showed that judgments of coreference are substantially degraded when a
pronoun c-commands its antecedent, as predicted by
the Principle C constraint. These results show that Principle C captures genuine acceptability contrasts, but
leave open the question of whether the constraint
impacts the real-time construction of coreference relations. The three on-line experiments all used the gender
mismatch paradigm (Aoshima et al., submitted for publication; Van Gompel & Liversedge, 2003) to test where
active dependency formation occurs. All three experiments found a gender mismatch effect in the no-constraint conditions and no gender mismatch effect in the
Principle C conditions. These findings are therefore consistent with the hypothesis that upon encountering a cataphoric pronoun the parser actively searches for an
antecedent for the pronoun, except in positions that
are excluded by grammatical constraints. This in turn
suggests that constraints on coreference apply during
the earliest stages of sentence processing.
The results of Experiment 1 showed the contrast
between the no-constraint conditions and the Principle
C conditions, and a fifth control condition demonstrated
that the gender mismatch effect in the no-constraint conditions could not simply be an effect of additional discourse
complexity. However, in this experiment an additional difference in structure predictability was confounded with
the difference in the relevance of Principle C. The lack
of a gender mismatch effect in the Principle C conditions
could have been due to the effects of Principle C, or it
may have reflected the fact that the critical antecedent
position could not be reliably predicted far in advance.
Experiments 2 and 3 were designed to match the conditions in terms of predictability. Experiment 2 contained
structures that shared equally strong indicators (prior to
405
the pronoun) of the upcoming critical subject position,
while in Experiment 3 the critical antecedent position
was equally unpredictable at the pronoun across the
two pairs of conditions. In both experiments we
obtained essentially the same results as in Experiment
1: a significant reading time slowdown associated with
gender mismatch in the no-constraint conditions, but
no corresponding slowdown in the Principle C conditions. These findings suggest that the parser considered
the critical subject noun phrase as a potential antecedent
in advance of bottom-up morphological information,
but only if the subject position was not ruled out by a
grammatical constraint on coreference.
These findings clearly parallel results from previous
work on filler-gap dependencies that has shown that
those dependencies are parsed using an active search
mechanism that posits gaps in advance of bottom-up
confirmation, but does not posit gaps in positions that
are excluded by island constraints on long-distance
dependencies (Phillips, 2006; Stowe, 1986; Traxler &
Pickering, 1996). Backwards anaphora constructions
have the same dependent-element-first property as filler-gap dependencies, but differ from filler-gap dependencies in a number of respects, most notably the fact that
potential antecedent noun phrases are overt and hence
easier to identify based on bottom-up information than
phonologically null gaps. The similarity in the processing mechanisms for these two different-looking types
of constructions raises the possibility that the parser
might draw on a single mechanism for all linguistic
dependencies that share this same basic dependent-element-first structure. It also suggests that the active
search mechanism is not a direct consequence of the fact
that gaps are difficult to identify using bottom-up information alone, since the same mechanism is found in the
search for antecedents of pronouns.
In light of the evidence from Experiment 3 for continuous (rather than instantaneous) resolution of cataphoric pronouns, the difference in the amplitude and timing
of the gender mismatch effect in Experiment 2 relative
to the other two experiments cannot be accounted in
terms of the weaker predictability of the second subject
position in that experiment. Instead, the strength of the
mismatch effect may be dependent on the distance of the
critical subject position from its structural predictor.
This idea was supported by the stronger mismatch effect
observed in Experiment 3. Although the critical subject
position in that experiment was entirely unpredictable
at the point when the pronoun was first processed, the
subsequent subordinator (e.g., while) made it possible
to predict the head of the subject noun phrase three
words before it was confirmed by bottom-up information. This increased distance may have enabled the
parser to fully process the subordinator and the accompanying structural prediction for a subject position.
Although it is possible that the parser did not initiate
406
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
processing of the pronoun-antecedent relation until bottom-up information about the antecedent was encountered, as assumed by Van Gompel and Liversedge
(2003), this would entail that the consequences of grammatical constraints such as Principle C are computed
faster than the gender of the antecedent can be identified, and would also make it more difficult to account
for the delayed gender mismatch effect in Experiment 2.
Generality of active search mechanism
To understand how the parser actively constructs
long-distance dependencies in advance of bottom-up
information it is important to know under what circumstances dependency formation can be initiated. In the
domain of wh-dependencies there is good reason for
the parser to initiate the search for a gap as soon as a
wh-phrase is encountered, since the wh-phrase itself provides sufficient evidence that a gap must occur somewhere in the sentence. Matters are less straightforward
in the case of backwards anaphora, since the occurrence
of a pronoun early in a sentence does not normally guarantee that an antecedent for the pronoun will appear later in the sentence. Furthermore, the occurrence of a
pronoun does not normally guarantee that even a potential antecedent position will appear. However, the results
of Experiment 3 indicate that active search can be initiated even when evidence for an upcoming potential
antecedent position does not occur until a few words
after the pronoun. Of course, we cannot at present
exclude the possibility that the generality of active
search in our studies was related to the fact that the target and filler items in our studies were designed to ensure
that all pronouns found an intra-sentential antecedent.
This aspect of the design was well-motivated, based on
our interest in the absence of gender mismatch effects
in the Principle C conditions, but it may have exaggerated the evidence for active search.
A related question about the generality of active
search involves the range of positions where the parser
considers potential antecedents to occur. In the case of
wh-constructions previous evidence indicates that gaps
are actively posited in object positions (Stowe, 1986)
and subject positions (Lee, 2004) of the clause that is
local to the wh-phrase, and also in embedded clauses
that could not have been anticipated at the point when
the wh-phrase was first encountered (Aoshima et al.,
2004; Frazier & Clifton, 1989; Phillips, 2006; Phillips,
Kazanina, & Abada, 2005). It should be noted that the
studies reported here share with previous studies on
backwards anaphora processing (Aoshima et al., submitted for publication; Van Gompel & Liversedge,
2003) the limitation that they show active construction
of anaphoric dependencies only in subject positions.
This limitation does not undermine our conclusions
about the effects of Principle C, but it leaves open the
question of whether the active search for pronoun antecedents is as general as the active search for gaps.
Temporal priority for structural information
As discussed in the Introduction, the existence of
gender mismatch effects implies that the parser constructs anaphoric dependencies before gender information about a potential antecedent noun phrase is fully
processed. Our current results add to this the finding
that structural constraints on coreference such as Principle C are implemented before gender information is processed, providing further evidence for a temporal
priority for structural information in the processing of
backwards anaphora. Van Gompel and Liversedge
(2003) argue that this timing advantage for syntactic
information provides evidence against ‘strongly interactive’ processing models and favors ‘modular’ models
that give priority to syntactic evidence (p. 138). We
would argue, however, that this conclusion is not
required, since the temporal priority for syntactic information in processing backwards anaphora can be
explained without the need to appeal to architectural
constraints that specifically delay the use of certain types
of information. We suggest that structural information
has priority in this domain because it is often the only
type of information about a potential antecedent that
is reliably derivable before bottom-up information
about the noun phrase is encountered. If the parser
encounters structural evidence for an upcoming clause,
then it can immediately and reliably predict that the
clause will have a subject noun phrase, and it could also
immediately evaluate whether that subject noun phrase
is in a position that makes it a structurally licit or illicit
antecedent for a previously encountered pronoun. This
information could all be computed before any bottomup information about the subject noun phrase is encountered. In contrast, other information about the subject
noun phrase such as its morphological features and its
semantic match to a previously encountered pronoun
cannot be evaluated except via bottom-up information,
because these properties are not structurally predictable.
Thus, under this view syntactic information plays a crucial role because it enables the parser to make predictions about upcoming material earlier than any other
type of information, and hence there is no need to
impose architectural constraints that force certain information types to have priority. Note that this line of reasoning might apply differently in languages that display
richer morphological agreement than English, such that
it may be possible to reliably predict morphological
properties of an upcoming noun in advance of the noun
itself.
Studies on the processing of forwards anaphora have
also explored the issue of priority for structural information, and the discussion has focused on the question of
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
whether structural constraints on coreference immediately restrict the parser’s selection of candidate antecedents for pronouns and reflexives. Some studies have
found immediate effects of structural constraints in processing forwards anaphora (e.g., Nicol & Swinney, 1989;
Sturt, 2003), while others have reported effects of
ungrammatical antecedents on the processing of pronouns (e.g., Badecker & Straub, 2002; Kennison, 2003;
see also Runner, Sussman, & Tanenhaus, 2003). Sturt
(2003) proposes a ‘binding-as-defeasible-filter’ model
that attributes the differing results across studies to different processing stages. Sturt suggests that syntactic
constraints on coreference are respected at an initial
phase of processing, but may be violated at later stages.
He observes time-course differences in his own studies in
the effects of structurally accessible and inaccessible
antecedents. Further, Sturt’s account may capture the
fact that in previous studies even highly salient structurally inaccessible antecedents have failed to elicit a gender
mismatch effect, a measure of early anaphor resolution
(for more detailed review see Kazanina, 2005). In the
respect that we observed no effects of structurally inaccessible antecedents in our studies, our results are compatible with Sturt’s account. Note, however, that care
should be taken in comparing the resolution of forwards
and backwards anaphora, due to differences in their
temporal dynamics and due to differences in the representations that must be searched for an antecedent. Forwards anaphora involves a retrospective search for an
antecedent in memory, and multiple candidate antecedents may be considered simultaneously. Backwards
anaphora, in contrast, involves a prospective search
for an antecedent. Consequently, potential antecedents
do not need to be retrieved from memory, and may be
evaluated in succession. In contrast to some studies of
forwards anaphora, our studies found no delayed effect
of structurally inaccessible antecedents. This difference
may reflect either the temporal dynamics of backwards
anaphora resolution, or the different representations
that are searched for an antecedent.
Analogous effects of giving priority to one type of
information over others during anaphora processing
have also been discussed by Sanford and Garrod
(1989) for cross-sentential forwards anaphor resolution.
Their findings suggest that bottom-up morphological
information about a noun phrase and a pronoun can
set in motion a process they call ‘bonding’: the formation of a tentative (and ultimately incorrect) dependency
in cases like ‘‘Harry was sailing to Ireland. It sank without trace’’ on the basis of the shared singular morphology of Ireland and it (originally reported in Sanford,
Garrod, Lucas, & Henderson, 1984). These findings
are similar to ours in the sense that one type of information has temporal priority over others in the processing
of anaphoric dependency. However, whereas the Sanford and Garrod’s effect relies on bottom-up morpho-
407
logical information to begin the resolution process, our
cases crucially depend on the combined use of information from the pronoun and structural expectations to
create a dependency in advance of the bottom-up information, although the full details are filled in only after
the bottom-up information about the antecedent is
encountered. Therefore, we think that these two types
of effects are probably driven by distinct mechanisms.
Implications for ungrammatical representations
Finally, our results are also relevant to questions
about the nature of grammatical constraints and
how they are implemented. In some linguistic frameworks there is a clear distinction between the range of
structures that the grammar generates and the range of
structures that satisfy all grammatical constraints or ‘filters.’ This overgenerate-and-filter property is explicitly
advocated in such frameworks as Optimality Theory
(Prince & Smolensky, 1993/2004) and Government and
Binding Theory (Chomsky, 1981). A different perspective on illicit representations is that ill-formedness is
the result of an inability to generate a structure, such
that there is a closer match between the set of wellformed structures and the set of representable structures.
This is more characteristic of analyses in such approaches as Head-Driven Phrase Structure Grammar (HPSG:
Pollard & Sag, 1994) and the Minimalist Program
(Chomsky, 1995). To the extent that parsing is seen as
reflective of the linguistic knowledge modeled by these
theories, our results are consistent with the latter view,
at least for Principle C. Speakers do not appear to generate representations that violate Principle C and then
reject them based on application of a later filter. Instead,
it appears that Principle C violations simply do not
occur to speakers of English. Note that this finding contrasts with the conclusion from an interesting but controversial set of findings by Freedman and Forster
(1985), who claimed based on the results of a series of
sentence-matching studies that the parser considers illicit
filler-gap dependencies that violate syntactic island constraints (for alternative interpretations see Crain &
Fodor, 1987; Stowe, 1992). Note, also, that our results
cannot exclude the possibility that the parser does temporarily construct representations that violate Principle
C, but successfully filters these representations before
any bottom-up information about an antecedent is
presented.
Conclusion
We have shown that upon encountering a cataphoric
pronoun the parser initiates an active search for an antecedent in the following material. Importantly, during
this search the parser does not consider positions that
408
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
are inaccessible antecedent positions due to a structural
constraint on coreference, Principle C. These results support a view whereby syntactic constraints immediately
restrict active search processes, whether the search is
for a thematic position, as in wh-dependencies, or for a
referent, as in backwards anaphora. More generally,
our results suggest that the mechanisms used for processing wh-dependencies and referential dependencies
like backwards anaphora might be largely identical.
Acknowledgments
We are grateful to Roger van Gompel and to an
anonymous reviewer for valuable feedback on an earlier
draft. This research was supported by grants to CP from
the National Science Foundation (#BCS-0196004) and
the Human Frontiers Science Program (#RGY-0134).
EFL was supported by an NSF Graduate Fellowship.
Appendix A. Supplementary data
Supplementary data associated with this article can be
found, in the online version, at doi:10.1016/j.jml.2006.09.003.
References
Aoshima, S. (2003). The grammar and parsing of wh-dependencies. PhD dissertation, University of Maryland, College
Park.
Aoshima, S., Phillips, C., & Weinberg, A. S. (2004). Processing
filler-gap dependencies in a head-final language. Journal of
Memory and Language, 51, 23–54.
Aoshima, S., Yoshida, M., & Phillips, C. (submitted for
publication). Incremental processing of coreference and
binding in Japanese.
Badecker, W., & Straub, K. (2002). The processing role of
structural constraints on the interpretation of pronouns and
anaphora. Journal of Experimental Psychology: Learning
Memory and Cognition, 28, 748–769.
Baker, M. A. (1991). On some subject/object non-asymmetries
in Mohawk. Natural Language and Linguistic Theory, 9,
537–576.
Bourdages, J. S. (1992). Parsing complex NPs in French. In H.
Goodluck & M. S. Rochemont (Eds.), Island constraints:
Theory, acquisition and processing (pp. 61–87). Netherlands:
Kluwer Academic Publishers.
Brody, M. (1994). Phrase structure and dependence. London:
Ms. University College.
Bruening, B. (2001). Syntax at the edge: cross-clausal phenomena and the syntax of Passamaquoddy. PhD dissertation,
MIT.
Chomsky, N. (1973). Conditions on transformations. In S.
Anderson & P. Kiparsky (Eds.), A festschrift for morris halle
(pp. 232–286). New York: Holt, Rinehart & Winston.
Chomsky, N. (1981). Lectures on government and binding.
Dordrecht: Foris.
Chomsky, N. (1995). The minimalist program. Cambridge,
Mass: MIT Press.
Clark, H. H. (1973). The language-as-fixed-effect fallacy: a
critique of language statistics in psychological research.
Journal of Verbal Learning and Verbal Behavior, 12,
335–359.
Clifton, C., Kennison, S. M., & Albrecht, J. E. (1997). Reading
the words him and her: implications for parsing principles
based on frequency and on structure. Journal of Memory
and Language, 36, 276–292.
Cowart, W., & Cairns, H. S. (1987). Evidence for an anaphoric
mechanism within syntactic processing: some reference
relations defy semantic and pragmatic constraints. Memory
& Cognition, 15, 318–331.
Crain, S., & Fodor, J. D. (1985). How can grammars help
parsers. In D. R. Dowty, L. Karttunen, & A. M. Zicky
(Eds.), Natural language parsing: psychological, computational, and theoretical perspectives (pp. 94–128). Cambridge:
Cambridge University Press.
Crain, S., & Fodor, J. D. (1987). Sentence matching and
overgeneration. Cognition, 26, 123–169.
Crain, S., & McKee, C. (1985). The acquisition of structural
restrictions on anaphora. In S. Berman, J. Choe, & J.
McDonough (Eds.), Proceedings of NELS15. Amherst, MA:
GLSA Publications.
Fodor, J. A. (1983). The modularity of mind. Cambridge, MA:
MIT Press.
Fodor, J. D. (1978). Parsing strategies and constraints on
transformations. Linguistic Inquiry, 9, 427–473.
Frazier, L. (1987). Sentence processing: a tutorial review. In M.
Coltheart (Ed.), Attention and performance XII. Hillsdale,
N.J.: Erlbaum.
Frazier, L., & Clifton, C. (1989). Successive cyclicity in the
grammar and the parser. Language and Cognitive Processes,
4, 93–126.
Freedman, S. E., & Forster, K. I. (1985). The psychological
status of overgenerated sentences. Cognition, 19, 101–131.
Garnsey, S. M., Tanenhaus, M. K., & Chapman, R. M. (1989).
Evoked potentials and the study of sentence comprehension.
Journal of Psycholinguistic Research, 18, 51–60.
Gordon, P. C., & Hendrick, R. (1997). Intuitive knowledge of
linguistic co-reference. Cognition, 62, 325–370.
Harris, C. L., & Bates, E. A. (2002). Clausal backgrounding
and pronominal reference: a functionalist approach to ccommand. Language and Cognitive Processes, 17,
237–269.
Heim, I. (1992). Anaphora and semantic interpretation: a
reinterpretation of Reinhart’s approach. In U. Sauerland &
O. Percus (Eds.), The interpretive tract (pp. 205–246).
Cambridge, MA: MIT Working Papers in Linguistics.
Hirst, W., & Brill, G. A. (1980). Contextual aspects of pronoun
assignment. Journal of Verbal Learning and Verbal Behavior,
19, 168–175.
Huang, C.-T. J. (1982). Move-wh in a language without whmovement. The Linguistic Review, 1, 369–416.
Jelinek, E. (1984). Empty categories, case, and configurationality. Natural Language and Linguistic Theory, 2, 39–76.
Just, M. A., Carpenter, P. A., & Woolley, J. D. (1982).
Paradigms and processes and in reading comprehension.
Journal of Experimental Psychology: General, 3,
228–238.
N. Kazanina et al. / Journal of Memory and Language 56 (2007) 384–409
Kaan, E., Harris, A., Gibson, E., & Holcomb, P. (2000). The
P600 as an index of syntactic integration difficulty. Language and Cognitive Processes, 15, 159–201.
Kazanina, N. (2005). The acquisition and processing of backwards anaphora. Ph.D. dissertation, University of Maryland, College Park.
Kazanina, N., & Phillips, C. (2001). Coreference in child russian:
distinguishing syntactic and discourse constraints. In A. Do,
L. Domiı́nguez, & A. Johansen (Eds.), Proceedings of the 25th
Annual Boston University Conference on Language Development (pp. 413–424). Somerville MA: Cascadilla Press.
Kennison, S. M. (2003). Comprehending the pronouns her, him,
and his: Implications for theories of referential processing.
Journal of Memory and Language, 49, 335–352.
Lee, M. (2004). Another look at the role of empty categories in
sentence processing (and grammar). Journal of Psycholinguistic Research, 33, 51–73.
Masson, M. E. J., & Loftus, G. R. (2003). Using confidence
intervals for graphically based data interpretation. Canadian
Journal of Experimental Psychology, 57, 203–220.
Nicol, J. L., Fodor, J. D., & Swinney, D. (1994). Using crossmodal lexical decision tasks to investigate sentence processing. Journal of Experimental Psychology: Learning, Memory
& Cognition, 20, 1229–1238.
Nicol, J., & Swinney, D. (1989). The role of structure in
coreference assignment during sentence comprehension.
Journal of Psycholinguistic Research, 18, 5–19.
Phillips, C. (2006). The real-time status of island phenomena.
Language, 82, 4.
Phillips, C., Kazanina, N., & Abada, S. H. (2005). ERP effects
of the processing of syntactic long-distance dependencies.
Cognitive Brain Research, 22, 407–428.
Phillips, C. & Wagers, M. (in press). Relating structure and
time in linguistics and psycholinguistics. In: G. Gaskell
(ed.), Oxford handbook of psycholinguistics. Oxford, UK:
Oxford University Press.
Pickering, M. J. (1993). Direct association and sentence
processing: a reply to Gorrell and to Gibson and Hickok.
Language and Cognitive Processes, 8, 163–196.
Pickering, M. J., Barton, S., & Shillcock, R. (1994). Unbounded
dependencies, island constraints and processing complexity.
In C. Clifton, L. Frazier, & K. Rayner (Eds.), Perspectives on
sentence processing (pp. 199–224). London: Lawrence
Erlbaum.
Pollard, C., & Sag, I. A. (1994). Head-driven phrase structure
grammar. Chicago: University of Chicago Press.
Prince, A., & Smolensky, P. (1993/2004). Optimality theory:
Constraint interaction in generative grammar. Technical
409
Report, Rutgers University and University of Colorado at
Boulder, 1993. Revised version published by Blackwell,
2004. Rutgers Optimality Archive 537.
Reinhart, T. (1983). Anaphora and semantic interpretation.
London: Croom Helm.
Reinhart, T., & Reuland, E. (1993). Reflexivity. Linguistic
Inquiry, 24, 657–720.
Resnik, P. & Elkiss, A. (2003). The linguist’s search engine:
Getting started guide. Technical Report LAMP-TR-108,
University of Maryland.
Ross, J. R. (1967). Constraints on variables in syntax. Ph.D.
dissertation, MIT.
Runner, J. T., Sussman, R. S., & Tanenhaus, M. K. (2003).
Assignment of reference to reflexives and pronouns in
picture noun phrases: evidence from eye movements.
Cognition, 89, B1–B13.
Sanford, A. J., Garrod, S. C., Lucas, A., & Henderson, R.
(1984). Pronouns without antecedents? Journal of Semantics, 2, 303–318.
Sanford, A. J., & Garrod, S. C. (1989). What, when and how?
Questions of immediacy in anaphoric reference resolution.
Language and Cognitive Processes, 4, 235–262.
Speas, M. (1990). Phrase structure in natural language. Dordrecht: Kluwer Academic Publishers.
Stowe, L. (1986). Evidence for online gap creation. Language
and Cognitive Processes, 1, 227–245.
Stowe, L. A. (1992). The processing implementation of syntactic constraints: the sentence matching debate. In H. Goodluck & M. Rochemont (Eds.), Island constraints: Theory,
acquisition and processing (pp. 419–443). Dordrecht: Kluwer
Academic Publishers.
Sturt, P. (2003). The time-course of the application of binding
constraints in reference resolution. Journal of Memory and
Language, 48, 542–562.
Sussman, R. S., & Sedivy, J. C. (2003). The time-course of
processing syntactic dependencies: evidence from eye-movements during spoken wh-questions. Language and Cognitive
Processes, 18, 143–163.
Traxler, M. J., & Pickering, M. J. (1996). Plausibility and
the processing of unbounded dependencies: an eye-tracking study. Journal of Memory and Language, 35,
454–475.
Van Gompel, R. P. G., & Liversedge, S. P. (2003). The influence
of morphological information on cataphoric pronoun
assignment. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 29, 128–139.
van Hoek, K. (1997). Anaphora and conceptual structure.
Chicago: University of Chicago Press.